[Paper Review] Fine-grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks

Yossi Adi, Einat Kermany+3 more|arXiv (Cornell University)|Aug 15, 2016

Topic Modeling27 references297 citations

TL;DR

The paper proposes a framework to analyze sentence embeddings by predicting low-level properties (length, word content, word order) from fixed-length representations, comparing CBOW, LSTM auto-encoders, and skip-thoughts.

ABSTRACT

There is a lot of research interest in encoding variable length sentences into fixed length vectors, in a way that preserves the sentence meanings. Two common methods include representations based on averaging word vectors, and representations based on the hidden states of recurrent neural networks such as LSTMs. The sentence vectors are used as features for subsequent machine learning tasks or for pre-training in the context of deep learning. However, not much is known about the properties that are encoded in these sentence representations and about the language information they capture. We propose a framework that facilitates better understanding of the encoded representations. We define prediction tasks around isolated aspects of sentence structure (namely sentence length, word content, and word order), and score representations by the ability to train a classifier to solve each prediction task when using the representation as input. We demonstrate the potential contribution of the approach by analyzing different sentence representation mechanisms. The analysis sheds light on the relative strengths of different sentence embedding methods with respect to these low level prediction tasks, and on the effect of the encoded vector's dimensionality on the resulting representations.

Motivation & Objective

• Develop a task-independent method to probe what information sentence embeddings encode.
• Compare different sentence representation methods (CBOW, LSTM auto-encoder, skip-thought) along low-level properties.
• Investigate how embedding dimensionality affects encoding of length, content, and order.

Proposed method

• Define three auxiliary prediction tasks (length, word-content, word-order) to evaluate sentence representations.
• Train simple classifiers on fixed-length representations to predict each task from s, w, and w1, w2.
• Use a large Wikipedia-derived corpus to generate balanced training, development, and test sets for each task.
• Compare CBOW (averaged word vectors) and encoder–decoder LSTM architectures across multiple embedding sizes.
• Include skip-thought vectors as an additional representation for comparison.
• Assess the impact of natural language properties by permuting word order and analyzing synthetic (random word) datasets.

Experimental results

Research questions

• RQ1 To what extent do different sentence representations encode sentence length, word content, and word order?
• RQ2 How does embedding dimensionality affect the encoding of length, content, and order across CBOW, LSTM auto-encoders, and skip-thoughts?
• RQ3 Do encoders rely on natural language word-order statistics, and how does permuting sentences affect encoding?
• RQ4 Is BLEU on reconstruction a reliable proxy for encoder quality on low-level encoding tasks?

Key findings

• LSTM auto-encoders effectively encode word order and content, with performance improving with dimensionality up to a point.
• CBOW surprisingly encodes length and word order information, and low-dimensional CBOW often outperforms more complex models on content tasks.
• Increasing dimensions benefits some tasks more than others; for LSTM encoders, too many hidden units can degrade word-content encoding.
• CBOW’s length encoding is partly due to embedding-norm decay with sentence length, and not solely due to word identity.
• Skip-thoughts preserve word-content and length similarly to high-dimensional LSTM encoders but rely more on natural language patterns for word order; permuting sentences markedly degrades its performance on some tasks.
• BLEU scores do not always reflect encoder quality for word content/word-order encoding.