News Topic Similarity Measure using Pretrained BERT Model

Utilizing Next Sentence Predictions. No Fine-tuning Required.

Ceshine Lee
Feb 10, 2019 · 6 min read

In this post we establish a topic similarity measure among the news articles collected from the New York Times RSS feeds. The main purpose is to familiarized ourselves with the (PyTorch) BERT implementation and pretrained model(s).

What is BERT?

We’ve seen transformers[4] used to train (unidirectional) language models in the OpenAI paper[3]. BERT uses a clever task design (masked language model) to enable training of bidirectional models, and also adds a next sentence prediction task to improve sentence-level understanding. As a result, BERT obtains new state-of-the-art results on eleven natural language processing tasks.

Here is a great series on BERT. If you’re already familiar with language models, you can start with Part 2 (otherwise you might want to check out Part 1):

We’ve also covered the OpenAI before in this post:

Problem Description

An example:

Title: What Is and Isn’t Affected by the Government Shutdown

Description: Transportation Security Administration officers checking passengers at Pittsburgh International Airport last week. The agency’s employees have called out sick in increased numbers across the country since the shutdown began.

Next Sentence Prediction

Firstly, we need to take a look at how BERT construct its input (in the pretraining stage). An input sequences consists of two “sentences”(explained below), each ends with a [SEP] token. And a [CLS] token is added to the head, whose corresponding hidden states are used to make the next sentence prediction.

To generate each training input sequence, we sample two spans of text from the corpus, which we refer to as “sentences” even though they are typically much longer than single sentences (but can be shorter also). [1]

The pretrained BERT model, according to the paper[1], achieves 97%~98% accuracy at this task with a balanced dataset.

Proposed Solution

Intuitively, if two news article covers the same topic, the title from one article and the description from another combined should also have a high next sentence score, since they are sentimentally similar.

This is the converted input sequence of the previous example:

[CLS] what is and isn ’ t affected by the government shut ##down [SEP] transportation security administration officers checking passengers at pittsburgh international airport last week . the agency ’ s employees have called out sick in increased numbers across the country since the shut ##down began .[SEP]

And this is the above input sequence with the second sentence replaced by the description from another article:

[CLS] what is and isn ’ t affected by the government shut ##down [SEP] gr ##udge ##s can be good . they are one habit that humans have evolved to keep ourselves from the pain of breakup ##s and also from eating mo ##zza ##rella sticks for every meal . [SEP]


For the correct pairs (the title and description came from the same article), only 2.5% of them were give a lower than 50% next sentence score by the pretrained model (BERT-base-uncased). 97.3% of them has a scores above 90%. The following is an example that was predicted negative:

Title: Meanwhile: For a Knife, Dagger, Sword, Machete or Zombie-Killer, Just Ask These Ladies

Description: Whitehead’s Cutlery in Butte, Mont., is 128 years old and will gladly sharpen scissors sold generations ago.

It’s more tricky to evaluate other pairs, since we have no labeled data regarding to the similarity between articles. We can only qualitatively examine some of the pairs.

As an example, I use the title from the previous shown government shutdown article and generate 2,719 input sequences with the descriptions from the whole corpus. Here are the top five matches (highest next sentence scores):

Top 5 Descriptions Matching this Title: What Is and Isn’t Affected by the Government Shutdown1. Transportation Security Administration agents are among the most visible federal employees affected by the government shutdown.2. Transportation Security Administration agents at Chicago Midway International Airport on Dec. 22, the first day of the government shutdown.3. Damage from Hurricane Michael lingers throughout Marianna, Fla. The government shutdown has made things worse for many residents.4. Scientists aboard a National Oceanic and Atmospheric Administration ship in San Juan, P.R. The administration has furloughed many workers because of the government shutdown.5. Major federal agencies affected by the shutdown include the Departments of Agriculture, Commerce, Homeland Security, Housing and Urban Development, the Interior, Justice, State, Transportation, and the Treasury and the Environmental Protection Agency.

Interestingly, the actual description from the article did not make top 5 nor top 10. But still, the predicted descriptions seem reasonably good.

Conclusions and Future Work

In fact, one of the downstream task BERT was evaluated upon is Quora Question Pairs[5], which asks the model to determine if two questions are asking the same thing. If we can manually label some data, the results might be even better.

Other more efficient way (in terms of computation required) way to measure document or sentence similarity is doing sentence embeddings. In contrast to RNN and its variants (e.g. LSTM and GRU), extracting sentence embeddings from transformer models is not as straight forward (see this discussion).

Source Code

The pretrained model parameters are from the official Tensorflow implementation from Google:

The Jupyter Notebook used for this post:

The notebook used some preprocessing functions from the example script for sentence classification fine-tuning.


  1. Howard, J., & Ruder, S. (2018). Universal Language Model Fine-tuning for Text Classification.
  2. Radford, A., & Salimans, T. (2018). Improving Language Understanding by Generative Pre-Training.
  3. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. (2017). Attention is all you need.
  4. Z. Chen, H. Zhang, X. Zhang, and L. Zhao. 2018. Quora question pairs.

2019/03/23 Update

Related Posts

(This post is also published on my personal blog.)


Towards human-centered AI.

Medium is an open platform where 170 million readers come to find insightful and dynamic thinking. Here, expert and undiscovered voices alike dive into the heart of any topic and bring new ideas to the surface. Learn more

Follow the writers, publications, and topics that matter to you, and you’ll see them on your homepage and in your inbox. Explore

If you have a story to tell, knowledge to share, or a perspective to offer — welcome home. It’s easy and free to post your thinking on any topic. Write on Medium

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store