Sentiment analysis : Machine-Learning approach

Following up on my earlier post, as the frequency-based models were not very accurate and a good rule-based model was very hard to elaborate, we implemented what we known to be state-of-the-art methods for sentiment analysis on short sentences and make a list of the pros and cons of these methods. We train all of them on a 10.000 sentences dataset. These sentences are classified as positive, neutral, and negative by human experts. We the benchmark the models on a hold out sample of 500 sentences.

Word representations in a vector space

Feature extraction

To build a deep-learning model for sentiment analysis, we first have to represent our sentences in a vector space. We studied frequency-based methods in a previous post. They represent a sentence either by a bag-of-words, which is a list of the words that appear in the sentence with their frequencies, or by a term frequency – inverse document frequency (tf-idf) vector where the word frequencies in our sentences are weighted with their frequencies in the entire corpus.

These methods are very useful for long texts. For example, we can describe very precisely a newspaper article or a book by its most frequent words. However, for very short sentences, it’s not accurate at all. First, because 10 words are not enough to aggregate. But also because the structure of the sentence is very important to analyze sentiment and tf-idf models hardly capture negations, amplifications, and concessions. For instance, “Very good food, but bad for service…” would have the same representation as “Bad for food, but very good service!”.

Word vectors

We represent our sentences with vectors that take into account both the words that appear and the semantic structure. A first way to do this is to represent every word with an n-feature vector, and to represent our sentence with a n*length matrix. We can for instance build a vector of the same size as the vocabulary (10.000 for instance), and to represent the i-th word with a 1 in the i-th position and 0 elsewhere.

Tomas Mikolov developed another way to represent words in a vector space, with features that capture the semantic compositionality. He trains the following neural network on a very large corpus:

Neural network trained to get Word2Vec's word vectors

He trains this model and represents the word “ants” by the output vector of the hidden layer. The features of these word vectors we obtain capture most of the semantic information, because it captures enough information to evaluate the statistical repartition of the word that follows “ants” in a sentence.

What we do is similar. We represent every word by an index vector. And we integrate in our deep learning model a hidden layer of linear neurons that transforms these big vectors into much smaller ones. We take these smaller vectors as an input of a convolutional neural network. We train the model as a whole, so that the word vectors we use are trained to fit the sentiment information of the words, i.e. so that the features we get capture enough information on the words to predict the sentiment of the sentence.

Sentence representations

Doc2vec

We want to build a representation of a sentence that takes into account not only the words that appear, but also the sentence’s semantic structure. The easiest way to do this is to superpose these word vectors and build a matrix that represents the sentence. There is another way to do it, that was also developed by Tomas Mikolov and is usually called Doc2Vec.

He modifies the neural network we used for Word2Vec, and takes as an input both the word vectors that come before, and a vector that depends on the sentence they are in. We will take the features of this word vector as parameters of our model and optimize them using a gradient descent. Doing that, we will have for every sentence a set of features that represent the structure of the sentence. These features capture most of the useful information on how the words follow each other.

Neural Network trained to get Doc2Vec's document vectors

Pros and cons for sentiment analysis

These document vectors are very useful for us, because the sentiment of a sentence can be deduced very precisely from these semantic features . As a matter of fact, users writing reviews with positive or negative sentiments will have completely different ways of composing the words. Feeding a logistic regression with these vectors and training the regression to predict sentiment is known to be one of the best methods for sentiment analysis, both for fine-grained (Very negative / Negative / Neutral / Positive / Very positive) and for more general Negative / Positive classification.

We implemented and benchmarked such a method but we chose not to productionalize it. As a matter of fact, building the document vector of a sentence is not an easy operation. For every sentence, we have to run a gradient descent in order to find the right coefficients for this vector. Compared to our other methods for sentiment analysis, where the preprocessing is a very short algorithm (a matter of milliseconds) and the evaluation is almost instantaneous, Doc2Vec classification requires a significant hardware investment and/or takes much longer to process. Before taking that leap, we decided to explore representing our sentences by a matrix of word vectors and to classify sentiments using a deep learning model.

Convolutional neural networks

Convolutional neural networks

The next method we explored for sentiment classification uses a multi-layer neural network with a convolutional layer, multiple dense layers of neurons with a sigmoid activation function, and additional layers designed to prevent overfitting. We explained how convolutional layers work in a previous article. It is a technique that was designed for computer vision, and that improves the accuracy of most image classification and object detection models.

The idea is to apply convolutions to the image with a set of filters, and to take the new images it produces as inputs of the next layer. Depending on the filter we apply, the output image will either capture the edges, or smooth it, or sharpen the key patterns. Training the filter’s coefficients will help our model build extremely relevant features to feed the next layers. These features work like local patches that learn compositionality. During the training, it will automatically learn the best patches depending on the classification problem we want to solve. The features it learns will be location-invariant. It will convolve exactly the same way an object that is at the bottom of the frame and an object that is at the top of the frame. This is key not only for object detection, but for sentiment analysis as well.

Convolution used for edge detection

Convolution used for edge detection

Applications in Natural Language Processing

As these models became more and more popular in computer vision, a lot of people tried to apply them in other fields. They had significantly good results in speech recognition and in natural language processing. In speech recognition, the trick is to build the frequency intensity distribution of the signal for every timestamp and to convolve these images.

For NLP tasks like sentiment analysis, we do something very similar. We build word vectors and convolve the image built by juxtaposing these vectors in order to build relevant features.

Intuitively, the filters will enable us to highlight the intensely positive or intensely negative words. They will enable us to understand the relation between negations and what follows, and things like that. It will capture relevant information about how the words follow each other. It will also learn particular words or n-grams that bear sentiment information. We then feed a fully connected deep neural network with the outputs of these convolutions. It selects the best of these features in order to classify the sentiment of the sentence. The results on our datasets are pretty good.

Convolutional neural networks for Natural Language Processing

LSTM

We also studied, implemented and benchmarked the Long Short-Term Memory Recurrent Neural Network model. It has a very interesting architecture to process natural language. It works exactly as we do. It reads the sentence from the first word to the last one. And it tries to figure out the sentiment after each step. For example, for the sentence “The food sucks, the wine was worse.”. It will read “The”, then “food”, then “sucks”, “the” and “wine”. It will keep in mind both a vector that represents what came before (memory) and a partial output. For instance, it will already think that the sentence is negative halfway through. Then it will continue to update as it processes more data.

Recurrent neural networks - The human way to do sentiment analysis

This is the general idea, but the implementation of these networks is much more complex because it is easy to keep recent information in mind, but very difficult to have a model that captures most of the useful long-term dependencies while avoiding the problems linked to vanishing gradient.

This RNN structure looks very accurate for sentiment analysis tasks. It performs well for speech recognition and for translation. However, it slows down the evaluation process considerably and doesn’t improve accuracy that much in our application so should be implemented with care.

Sentiment trees – RNTN model

Richard Socher et al. describe in the paper Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank another cool method for sentiment analysis. He says that every word has a sentiment meaning. The structure of the sentence should enable us to compose these sentiments in order to get the overall sentiment of the sentence.

Stanford Sentiment Treebank example

They implement a model called the RNTN. It represents the words by vectors and takes a class of tensor-multiplication-based mathematical functions to describe compositionality. Stanford has a very large corpus of movie reviews turned into trees by their NLP libraries. Every node is classified from very negative to very positive by a human annotator. They trained the RNTN model on this corpus, and got very good results. Unfortunately, they train it on IMDB movie reviews data. But it doesn’t perform quite as well on our reviews.

The big advantage of this model is that it is very interpretable. We can understand very precisely how it works. We can visualize which words it detects to be positive or negative, and how it understands the compositions. However, we need to build an extremely large training set (around 10.000 sentences with fine-grain annotations on every node) for every specific application. As we continue to gather more and more detailed training data, this is just one of the types of models we are exploring to continue improving the sentiment models we have in production!


Semantic Convolution with Word2Vec

At Reputation.com, we work with millions of online reviews from hundreds of sources. One of the unusual characteristics of reviews compared to the vast majority of text corpora is that, almost by definition, reviews are structured in such a way that they can be categorized (in one or many dimensions depending on the review site and/or industry). However, we often find ourselves doing text classification/tagging on topics that are not already labeled by the review site.  This article is an informal introduction to a set of techniques we have developed to leverage existing unlabeled corpora in conjunction with the labeled data. In particular, we present a semi-supervised learning algorithm for multi-label text classification.

In recent years, a lot of text classification projects have used supervised learning methods (Naive Bayes, SVM) primarily due to their substantial improvements over non-supervised strategies such as traditional clustering in NLP tasks. Until very recently, most NLP classification work was done with the traditional Bag of Words (BOW) approach – perhaps with a bit of context through the use of a limited range of N-grams and skip-grams. BOW is a feature extraction technique where the text is represented as the frequency of each word in the document, disregarding grammar and ordering but keeping multiplicity. In most cases, defining a pipeline combining the BOW feature extraction technique with a Tf-Idf transform and a simple classifier (Naive Bayes, SVM) produces decent results with respect to most classification metrics.

Semi-Supervised Learning with Word2Vec

In most tutorials, Word2Vec is presented as a stand-alone neural net preprocessor for feature extraction. Word2Vec generates a vector for each word in the text corpora in higher-dimensional space such that words that share contextual meaning are located in close proximity to one another. To use Word2Vec for classification, each word can be replaced by its corresponding word vector and usually combined through a naive algorithm such as addition with normalization or cross product to get a sentence or text vector. Then, using these document vectors we could use a simple classifier for multi-label classification. The advantage of using Word2Vec over a simple BOW feature extraction technique is it supports semi-supervised learning, since the vocabulary from the labeled and unlabeled text can be used to generate the word vectors. This allows the words to have more contextual meaning. However, we have found that this approach does not appear to provide significant improvements over a BOW approach especially when there isn’t a lot of labeled data for training the classifier.

Semantic Convolution for Low Support Topics

A common problem that is seen in multi-label text classification is a major imbalance of labels in a textual corpora. We often see cases where most (>60%) of the sampled data is about the most prevalent topic, and more than half the topic labels exist in <0.1% of the sampled data. Almost inherently with NLP and a BOW approach, this causes a p (number of features) >> n (size of training corpus) problem. Based on a general rules of thumb, getting 1,000 training examples for the low support topic would require millions of labeled training examples, which is prohibitively expensive.

In this world of ‘big data’ the data itself is actually cheap, but developing a tagged training set can be expensive. In the course of our development, we devised an elegant and scalable way to develop and maintain a robust training set across tens of industries (this will be the topic of a separate blog post).

The premise of Semantic Convolution is simple: if a particular word is a good indicator of a particular label, then words with similar meanings (semantics) should also be good indicators of the label. Since we have qualitative evidence that Word2Vec vectors encode a semantic meaning, we can use it to help find words with similar meanings from non-labeled corpora. This allows us to apply a Semantic transform after getting the term frequencies in the BOW pipeline, and before applying the Tf-Idf transform. To apply the Semantic Transform, we use the Word2Vec data to generate a correlation matrix between words with similar contextual meaning in the vocabulary.

Given vocabulary is a dictionary mapping each term with an index, the code to generate the correlation_matrix is:

correlation_matrix = scipy.sparse.identity(len(vocabulary), format="dok")
for idx, word in enumerate(vocabulary.keys()):
    similar_words = []

    try:
        similar_words = [x[0] for x in word2vec_model.most_similar(word, topn=5) if x[1] > 0.5]
    except:
        raise

    for similar_word in similar_words:
        if similar_word in vocabulary:
            correlation_matrix[vocabulary[word], vocabulary[similar_word]] = 1

Using this correlation matrix we can generate the term-document matrix with the augmented term frequencies.

term_frequency_vector += term_frequency_vector * correlation_matrix

Applying this transformation with the correlation matrix increases the word count of all words with contextually similar meaning in the text. This improves the feature collection for low support topics, which allows more precise classification of reviews about low support topics with higher confidence. This allows small amounts of labeled data to be more useful for the machine-learning model, which reduces the cost of developing a robust training set. Also, as mentioned above, this leverages semi-supervised learning from the unlabeled data by building the vocabulary and Word2Vec vectors based on the entire text corpora.

Conclusions

Ultimately the Semantic convolution provides more value from the little labeled data, and improves the performance of the machine leaning algorithm for classification tasks, especially the low support categories. Also, semi-supervised learning with Word2Vec leverages the information gained from the vast amounts of unlabeled data while increase both the precision and the support of the machine-learning model.

Dweep Shah and Anthony Johnson


Review Classification with Neural Network Models

Introduction

Extracting meaning from online reviews is key to turn seemingly anecdotal reviews into actionable customer satisfaction insights that point to improvement opportunities or authentic, and potentially differentiating strengths. One way to do that is to apply machine learning to automatically read customer reviews and identify the most relevant topics that are the subject of the review. With this information, you can find themes in what customers are saying about a business across thousands of reviews and then help businesses identify areas in which they are receiving a disproportionate number of negative reviews so that they can focus operational efforts on these areas and improve customer experience as well as their online reputation.

We have been working for a while on several approaches, models, and data sets to extract topics and categories from customer reviews with a high precision. In this post I will give an overview of a few neural network models that provide satisfactory results for physician-related reviews. To start, we built a taxonomy of categories that are relevant to physician reviews looking both at clinical patient experience topics from standard patient assessment surveys designed by CMS (Center for Medicare and Medicaid Services) as well as non-clinical topics related to parking, technology/amenities, and cleanliness that are commonly referred to in physician reviews. Then we gathered training data by having a group of crowd-sourced individuals tag a set of 10,000 reviews with the following categories (this is the subject of an upcoming blog entry):

  • Administrative Process
  • Bedside Manner
  • Cleanliness
  • Competence
  • Getting an Appointment
  • Likely/Unlikely to recommend
  • Parking
  • Responsiveness
  • Staff Courtesy
  • Technology/Amenities
  • Price/Billing issues
  • Wait Time

Given this training data, we used a biologically-inspired variant of Artificial Neural Networks to build a classifier that automatically assigns categories to online physician reviews. These neural network classifiers are based on how an animal’s visual cortex processes and exploits the strong spatially local correlation present in natural images. Those models are generally used for image recognition, but are being increasingly used in other fields, especially text classification. Given the promising results documented in this space, we decided to evaluate Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) with respect to our classification problem.

After an initial trial, we decided to focus our implementation on CNN models as they execute faster, are easier to understand, and had comparable results to RNN.

Principle of CNNs

screen-shot-2016-10-03-at-1-51-58-pmThe starting point of our CNNs was to represent a review by a matrix where each row is a vector that represents a word. This vector could be low-dimensional representations or one-hot vectors that index words into a vocabulary. Given this vector, you can then apply several convolutional filters on groups of rows followed by a 1-max pooling (the largest number from each feature map is recorded) in order to extract the meaning of the group of words considered at the beginning. Finally, a softmax layer is applied to generate assessed probabilities of the review belonging to each class.

CNN Implementation Approach

We generated a CNN model for each category independently broken into two classes: reviews that belong to this category and reviews that do not.

Thus, to produce all of the hidden parameters of these models, we fed them with reviews from the training data that were already categorized and modified the parameters incrementally in order to minimize a loss function (a function that represents the difference between the prediction and the real categories).

CNN Model Effectiveness

To assess the performance of these models, we split the 10,000 reviews into 8,000 reviews for training and 2,000 for testing. Given a model built on the training data, we predicted whether each review in the test data belonged in each category and assessed the precision and recall of our predictions with respect to each category.

For the largest categories, we found that our models delivered an overall precision of 81% and an overall recall of 75%. At first sight, those results did not appear very good. However, when we dug deeper, we found that when we considered at each element of crowd-sourced (human-based) training data to be a prediction in itself, these tags exhibited precision and recall metrics lower than 70% (versus the consensus of the group). Thus, our model outperformed human classification.

Furthermore, looking deeper into the training data, we realized that some reviews were truly ambiguous and the categories not precise or discerning enough, which resulted in a high degree a disagreement between humans evaluating the same review. After removing the most ambiguous reviews from the training data set, we observed a marked increase of the overall accuracy of the model. What to do about those ambiguous reviews or how to fine-tune the categories will be the subject of a future post.

Next Steps

The results from CNN models are promising, and we are pushing them further by experimenting with several modifications of the model such as: oversampling the training set in order to have balanced data for each category, splitting reviews by characters instead of by words, and initializing with a low-dimensional representation of words using Word2Vec. Stay tuned for further updates.


Meaningless Words to Useful Phrases in Spark – word2phrase

Introduction to word2phrase

When we communicate, we often know that individual words in the correct placements can change the meaning of what we’re trying to say.  Add “very” in front of an adjective and you place more emphasis on the adjective.  Add “york” in after the word “new” and you get a location.  Throw in “times” after that and now it’s a newspaper.

It follows that when working with data, these meanings should be known.  The three separate words “new”, “york”, and “times” are very different than “New York Times” as one phrase.  This is where the word2phrase algorithm comes into play.

Words to Phrases

At its core, word2phrase takes in a sentence of individual words and potentially turns bigrams (two consecutive words) into a phrase by joining the two words together with a symbol (underscore in our case).  Whether or not a bigram is turned into a phrase is determined by the training set and parameters set by the user.  Note that every two consecutive words are considered, so in a sentence with w1 w2 w3, bigrams would be w1w2, w2w3.

In our word2phrase implementation in Spark (and done similarly in Gensim), there are two distinct steps; a training (estimator) step and application (transform) step.

*For clarity, note that “new york” is a bigram, while “new_york” is a phrase.

Estimator Step

The training step is where we pass in a training set to the word2phrase estimator.  The estimator takes this dataset and produces a model using the algorithm.  The model is called the transformer, which we pass in datasets that we want to transform, i.e. sentences that with bigrams that we may want to transform to phrases.

In the training set, the dataset is an array of sentences.  The algorithm will take these sentences and apply the following formula to give a score to each bigram:

score(wi, wj) = (count(wiwj) – delta) / (count(wi) * count(wj))

where wi and wj are word i and word j, and delta is discounting coefficient that can be set to prevent phrases consisting of infrequent words to be formed.  So wiwj is when word j follows word i.

After the score for each bigram is calculated, those above a set threshold (this value can be changed by the user) will be transformed into phrases.  The model produces by the estimator step is thus an array of bigrams; the ones that should be turned to phrases.

Transformer Step

The transform step is incredibly simple; pass in any array of sentences to your model and it will search for matching bigrams.  All matching bigrams in the array you passed in will then be turned to phrases.

You can repeat these steps to produce trigrams (i.e. three words into a phrase).  For example, with “I read the New York Times” may produce “I read the new_york Times” after the first run, but run it again to get “I read the new_york_times”, because in the second run “new_york” is also an individual word now.

Example

First we create our training dataset; it’s a dataframe where the occurrences “new york” and “test drive” appears frequently.  (The sentences make no sense as they are randomly generated words.  See below for link to full dataframe.)

You can copy/paste this into your spark shell to test it, so long as you have the word2phrase algorithm included (available as a maven package with coordinates com.reputation.spark:word2phrase:1.0.1).

Download the package, create our test dataframe:

spark-shell –packages com.reputation.spark.word2phrase.1.0.1

import org.apache.spark.ml.feature.Word2Phrase

val wordDataFrame = sqlContext.createDataFrame(Seq(
(0, “new york test drive cool york how always learn media new york .”),
(1, “online york new york learn to media cool time .”),
(2, “media play how cool times play .”),
(3, “code to to code york to loaded times media .”),
(4, “play awesome to york .”),
.
.
.
(1099, “work please ideone how awesome times .”),
(1100, “play how play awesome to new york york awesome use new york work please loaded      always like .”),
(1101, “learn like I media online new york .”),
(1102, “media follow learn code code there to york times .”),
(1103, “cool use play work please york cool new york how follow .”),
(1104, “awesome how loaded media use us cool new york online code judge ideone like .”),
(1105, “judge media times time ideone new york new york time us fun .”),
(1106, “new york to time there media time fun there new like media time time .”),
(1107, “awesome to new times learn cool code play how to work please to learn to .”),
(1108, “there work please online new york how to play play judge how always work please .”),
(1109, “fun ideone to play loaded like how .”),
(1110, “fun york test drive awesome play times ideone new us media like follow .”)
)).toDF(“label”, “inputWords”)

We set the input and output column names and create the model (the estimator step, represented by the fit(wordDataFrame) function).

scala> val t = new Word2Phrase().setInputCol(“inputWords”).setOutputCol(“out”)
t: org.apache.spark.ml.feature.Word2Phrase = deltathresholdScal_f07fb0d91c1f

scala> val model = t.fit(wordDataFrame)

Here are some of the scores (Table 1) calculated by the algorithm before removing those below the threshold (note all the scores above the threshold are shown here).  The default values have delta -> 100, threshold -> 0.00001, and minWords -> 0.

Table 1
bigram score
 test drive  0.002214815139686856
 work please 0.002047826661381…
 new york  5.946183949006843E-4
 york new  -1.64600247723372…
 york york  -6.43001404062082…
 york how  -6.64999302561707…
 how new  -6.80666229773923…
 new new  -7.42968903739342…
 to new  -7.52757602015383E-5
 york to  -9.25567252744992…

only showing top 10 rows

So our model produces three bigrams that will be searched for in the transform step:

test drive
work please
new york

We then use this model to transform our original dataframe sentences and view the results.  Unfortunately you can’t see the entire row in the spark-shell, but in the out column it’s clear that all instances of “new york” and “test drive” have been transformed into “new_york” and “test_drive”.

scala> val bi_gram_data = model.transform(wordDataFrame)
bi_gram_data: org.apache.spark.sql.DataFrame = [label: int, inputWords: string … 1 more field]

scala> bi_gram_data.show()

Table 2
label inputWords out
0 new york test dri…  new_york test_dri…
1 online york new y…  online york new_y…
2 media play how co…  media play how co…
3 code to to code y…  code to to code y…
4 play awesome to y…  play awesome to y…
5  like I I always .   like I I always .
6 how to there lear…  how to there lear…
7 judge time us pla…  judge time us pla…
8 judge test drive …  judge test_drive …
9 judge follow fun …  judge follow fun …
 10  how I follow ideo…  how I follow ideo…
 11  use use learn I t…  use use learn I t…
 12  us new york alway…  us new_york alway…
 13  there always how …  there always how …
 14  always time media…  always time media…
 15 how test drive to…  how test_drive to…
 16  cool us online ti…  cool us online ti…
 17 follow time aweso…  follow time aweso…
 18  us york test driv…  us york test_driv…
 19  use fun new york …  use fun new_york …

only showing top 20 rows

The algorithm and test dataset (testSentences.scala) are available at this repository.


Natural Language Processing with Spark

One of the goals of the Analytics team has been to provide newer, more in-depth ways to analyze the millions of comments that Reputation aggregates from various sources for each customer. One way to do this is through natural language processing (NLP) techniques like part-of-speech(POS) tagging, named entity recognition(NER), and stemming/lemmatization. Combining these NLP techniques with our existing segmentation tools allows us to begin comparing statistics across sets defined by the language content of those comments. For example, we could look at the set of Walgreens comments that mention Rite-Aid and see that these had higher than average ratings in comparison to the total set of Walgreens comments.

These evaluations, however, initially required us to load the set of comments that we wished to analyze into Python, then run each comment through a natural language parser one at a time locally each time we wanted to run an analysis. The overhead required to parse each of these reviews began to impede our ability to rapidly test different types of analyses, so we began to look into alternative methods for achieving this goal. What we were ultimately looking for was a pre-processed database that would allow us to look up a comment by id and receive a set of POS tags, named entities, and lemmas without having to re-parse each comment each time. This natural language pre-processing would need to be done retroactively to the tens of millions of comments already stored in our database, as well as incrementally on any new comments that have been pulled in every few days.

Since much of our analysis framework was already implemented in Python, we began adding this new NLP piece in Python as well. Of the various NLP libraries available to Python at the time of this writing, the one that seemed to work best on the 2-3 sentence reviews in our database was the CoreNLP library from Stanford. Essentially CoreNLP comes with a series of models that have been trained on a large corpus of sample words for different languages (presently English, Arabic, Chinese, French and German). These models are then used to evaluate the likely part-of-speech of new inputs based on patterns learned from the original training data. The library also uses similar processes to determine which words in a given input are references to some named entity (for example an organization, individual name, or location name) and to identify the stem form of each word for easier pattern analysis.

The downside of using CoreNLP, however, is that in order to run, it starts up a new, separate Java process which is then passed one comment at a time for parsing. Starting up this Java process creates 5-10 minutes of overhead for processing a set of comments of any size, and even once this separate process is running it can take a few minutes to fully parse an average length comment (3-5 sentences). Thus to run all the millions of historical comments through CoreNLP in a serial fashion would be computationally infeasible. Instead, we decided to use Apache Spark to bring up a distributed cluster to run these comments through CoreNLP in parallel.

Spark provides a set of libraries in either Python, Scala, R, or Java that handle the hassle of creating a distributed cluster of nodes and efficiently distributing data between them. While it can be used for a wide variety of purposes, we used it to take the set of comments that we needed to evaluate and figure out how to split those comments amongst clusters of varying sizes in order to reduce the time necessary to run all of our historical data through CoreNLP. Using Spark also provided the added bonus of easily integrating with AWS’ Elastic Map-Reduce (EMR) service, which has an easy-to-use command line interface for bringing up clusters of EC2 nodes. Amazon has preconfigured settings to automatically pass the relevant information about each EMR cluster through to Spark so that we can easily bring up any number of nodes with the same code. This makes it easy to setup a cron task to automatically parse the last few days worth of reviews on a regular basis.

Additionally, while we originally set out to create a Python application to interact with Spark and CoreNLP, we eventually discovered that we needed the ability to more carefully control which information CoreNLP passed to each Spark process. Since Spark is capable of running multiple threads on each node in order to better parallelize and since each thread runs a separate version of our Spark application, we noticed that each Python application in each thread was instantiating its own CoreNLP Java process. This meant that if we had 4 threads running on the same node, we would also have 4 CoreNLP Java processes running on that node, which would slow that node’s performance to a crawl. To get around this, we had to translate our application into Scala instead. Scala allows for the existence of transient variables, which allowed us to write our code in such a way that when multiple threads are running on the same node, they all use the same CoreNLP Java process, but whenever a new node is brought up it brings up a new process. (Thanks to Databrick’s Spark/CoreNLP wrapper for this idea!)

Below is some of the code from our Scala-based Spark application. It is designed to do the following:

  1.     Pull in some number of reviews from our Vertica database.
  2.     Distribute those reviews to a cluster of independent nodes.
  3.     Run each review through the CoreNLP process for that node.
  4.     Format CoreNLP’s output so that it can uploaded back into Vertica
  5.     Upload the natural language data (POS tags, NER tags, and lemmas) back into the database

Click here for Github Gist

Once our Spark application was working  on local developer machines, we began testing running it through EMR’s distributed clusters instead. Initially we ran into some headaches getting Spark to fully utilize the resources made available to it through EMR. There is a line in the code above that talks about pulling in the number of nodes available through the Spark Config (val num_exec = sc.getConf.get(“spark.executor.instances”).toInt). This line tells spark how many nodes it has available so that it can partition the data accordingly. Below are two screenshots of the CPU usage per node in AWS from before this change and after it:

Screen Shot 2015-11-10 at 3.59.15 PM

Before proper partitioning – Notice that in this case, the node in blue is the only one that appears to be actually doing any parsing. This is because Spark defaults to assuming a single data partition, so it runs all the comments through the master node.

Screen Shot 2015-12-31 at 12.07.02 PM

After proper partitioning – By explicitly telling Spark how many nodes to use, we can see that it now runs some comments through all 8 nodes. (Thanks to Cloudera for explaining this and more about how to properly tune Spark jobs!)

 

Additionally we ran into some trouble getting EMR to communicate with Vertica through the database’s security restrictions, which involved playing with our VPN settings. Once these hurdles were dealt with though, we were able to begin testing the scaling power of this CoreNLP/Spark/EMR solution. The following graph shows the number of minutes it took Spark to run as dependent on the number of thousands of comments run per each instance in the EMR cluster. As you can see, the time to run increases linearly as a function of how many comments each node is required to run.

minutes_per_comments_per_node

Minute to Run vs. # of thousands of comments per node in cluster – This graph shows the time it takes Spark to run our process as a function of number of comments per each distributed node in the cluster. It shows a linear relationship more or less up until the point where there are more than a million comments per node.

The outlier point at 1000 on the x-axis (= 1 million comments per node) is from when we ran all of our historical comments. Further research is required to figure out why performance seems to have degraded for that point.

Interestingly, we also found that it seems when the number of comments per node increases above a about a million or so, the EMR task would fail without outputting any errors in the logs (this is what happened with the rightmost datapoint on the above graph). This may be due to insufficient resources to run the number of comments assigned to that node(we used Amazon’s m3.xlarge instances for each node on each run), but we haven’t done enough analysis to confirm this. The short-term solution to this problem was simply to provide more nodes and get the ratio of comments per node back down to around 1 million or so.