Is Neural Topic Modelling Better than Clustering? An Empirical Study on Clustering with Contextual Embeddings for Topics

Introduction

Topic modelling is an unsupervised method to uncover latent semantic themes among documents (Boyd-Graber et al., 2017). Neural topic models (NTMs) (Miao et al., 2016; Srivastava and Sutton, 2017) incorporating neural components have significantly advanced the modelling results than the traditional Latent Dirichlet Allocation (LDA; Blei et al. 2003). Later, contextualized word and sentence embeddings produced by pre-trained language models such as BERT (Devlin et al., 2019) have demonstrated the state-of-the-art results in multiple Natural Language Processing (NLP) tasks (Xia et al., 2020), which attracts attentions from the topic modelling community. Recent work has successfully incorporated these contextualized embeddings into NTMs, showing improved topic coherence than conventional NTMs that use Bag-of-Words (BoW) as document representations (Bianchi et al., 2021a, b; Jin et al., 2021). Despite the promising performance, existing NTMs are generally based on a variational autoencoder framework (VAE; Kingma and Welling 2013), which suffers from hyper-parameters tuning and computational overheads (Zhao et al., 2021). Moreover, the integration of the pre-trained embeddings to the standard VAE framework adds additional model complexity. With high-quality contextualized document representations, do we really need sophisticated NTMs to obtain coherent and interpretable topics?

Recent work (Aharoni and Goldberg, 2020; Sia et al., 2020; Thompson and Mimno, 2020; Grootendorst, 2020) has shown that directly congregating contextualized embeddings can get semantically similar word or document clusters. Specifically, Sia et al. (2020) cluster vocabulary-level word embeddings and obtain top words from each cluster using weighing and re-ranking, while Thompson and Mimno (2020) consider polysemy and perform token-level clustering. However, the use of term frequency (TF) to select topic words fails to capture the semantics of clusters precisely because words with high frequency may be common across different clusters. Grootendorst (2020) propose a class-based Term Frequency- Inverse Document Frequenc (c-TF-IDF) method that extract important words from each clustered documents, which tends to choose representative words within each cluster to form topics. However, it overlooks the global semantics between clusters which could be incorporated. In addition, all above works only compare the performance with the traditional LDA while ignoring the promising NTMs proposed recently. The performance of the clustering-based topic models is still yet uncovered.

Is neural topic modelling better than simple embedding clustering? This work compares the performance of NTMs and contextualized embedding-based clustering systematically. Our main focus is to provide insights by comparing the two paradigms for topic models, which has not been investigated before. We employ a straightforward framework for clustering. In addition, we explore different strategies to select topic words for clusters. We evaluate our approach on three datasets with various text lengths.

Our contributions are as follows: First, we find that directly clustering high-quality sentence embeddings can generate as good topics as NTMs, providing a simple and efficient solution to uncover latent topics among documents. Second, we propose a new topic word selecting method, which is the key to producing highly coherent and diverse topics. Third, we show that the clustering-based model is robust to the length of documents and the number of topics. Reducing the embedding dimensionality negligibly affects the performance but saves runtime. From our best knowledge, we are the first to compare with NTMs, using contextualized embeddings that produced by various transformer-based models.

Models

This study compares embedding clustering-based models with LDA and a series of existing NTMs as follows. Implementation details are supplied in Appendix A.

LDA (Blei et al., 2003): the representative traditional topic model in history that generates topics via document-topics and topic-words distributions.

ProdLDA (Srivastava and Sutton, 2017): a prominent NTM that employs the VAE (Kingma and Welling, 2013) to reconstruct the BoW representation.

CombinedTM (Bianchi et al., 2021a): extends ProdLDA by concatenating the contextualized SBERT (Reimers and Gurevych, 2019) embeddings with the original BoW as the new input to feed into the VAE framework.

ZeroShotTM (Bianchi et al., 2021b): also builds upon ProdLDA, but it replaces the original BoW with SBERT embeddings entirely.

BERT+KM (Sia et al., 2020): a clustering-based method that first uses K-Means (KM) to cluster word embeddings, then apply TF to weight and re-rank words to obtain topic words.

BERT+UMAP+HDBSCAN (i.e., BERTopic) (Grootendorst, 2020): a clustering-based method that first leverages HDBSCAN (McInnes and Healy, 2017) to cluster BERT embeddings of the sentences and Uniform Manifold Approximation Projection (UMAP) (McInnes et al., 2018) to reduce embedding dimensions, then use a class-based TFIDF (i.e. c-TF-IDF) to select topic words within each cluster. Note that BERTopic may not generate the specified number of topics.

Contextual Embeddings+UMAP+KM (our method CETopic): we use a simple clustering framework with contextualized embeddings for topic modelling, as shown in Figure 1. We first encode pre-processed documents to obtain contextualized sentence embeddings through pre-trained language models. After that, we lower the dimension of the embeddings before applying clustering methods (e.g., K-Means; KM) to group similar documents. Each cluster will be regarded as a topic. Finally, we adopt a weighting method to select representative words as topics.

We believe that high-quality document embeddings are critical for clustering-based topic modelling. We thus experiment with different embeddings including BERT, RoBERTa (Liu et al., 2019), and SBERT. We also adopt SimCSE (Gao et al., 2021), a recently proposed sentence embeddings of contrastive learning, that has shown the state-of-the-art performance on multiple semantic textual similarity tasks. Both supervised and unsupervised SimCSE are investigated in our experiment (e.g., Table 2).

Pre-trained contextualized sentence embeddings often have high dimensionalities. To reduce the computational cost, we apply the UMAP in our implementation to reduce the dimensionality while maintaining the essential information of the embeddings. We find that reducing dimensionality before clustering has a negligible impact on performance (Section 4.4).

We cluster the dimension-reduced sentence embeddings using K-Means because of its efficiency and simplicity. Semantically close documents are gathered together, and each cluster is supposed to represent a topic.

Topic Words for Clusters

Once we have a group of clustered documents, selecting representative topic words is vital to identify semantics of topics. Inspired by previous works (Ramos et al., 2003; Grootendorst, 2020), we explore several weighting metrics to obtain topic words in clusters. Let $n_{t,d}$ be the frequency of word $t$ in document $d$, $\sum_{t^{\prime}}n_{t^{\prime},d}$ be the total words’ frequency in the document, and $D$ be the entire corpus. Term Frequency-Inverse Document Frequency (TFIDF) is defined as $\mathbf{TFIDF}=\frac{n_{t,d}}{\sum_{t^{\prime}}n_{t^{\prime},d}}\cdot\log\left(\frac{|D|}{|\{d\in D:t\in d\}|}\right)$. While capturing the word importance across the entire corpus, TFIDF ignores that semantically similar documents have been grouped together. To address this issue, we consider two alternative strategies. First, we concatenate the documents within a cluster to be a single long document and calculate the term frequency of each word in each cluster:

where $n_{t,i}$ is the frequency of word $t$ in cluster $i$, $\sum_{t^{\prime}}n_{t^{\prime},i}$ is the total word frequency in the cluster. Second, for each cluster $i$, we apply TFIDF:

where ${n_{t,d}}_{i}$ denotes the frequency of word $t$ in document $d$, which is in cluster $i$, and $|D_{i}|$ is the number of documents in cluster $i$.

Besides the two local cluster-based strategies, we further incorporate the global word importance with local term frequency within each cluster:

and we combine the global word importance with term frequency across clusters:

where $|K|$ is the number of clusters and $|\{t\in K\}|$ is the number of clusters that word $t$ appears.

Experiments

We adopt three datasets of various text lengths in our experiments, namely 20Newsgroupshttp://qwone.com/~jason/20Newsgroups/, M10 (Lim and Buntine, 2015), and BBC News (Greene and Cunningham, 2006). We follow OCTIS (Terragni et al., 2021) to pre-process these raw datasets. The statistics of the datasets are shown in Table 1.

2 Evaluation Metrics

We evaluate the topic quality in terms of both topic diversity and topic coherence: Topic Diversity (TU) (Nan et al., 2019) measures the uniqueness of the words across all topics; Normalized Pointwise Mutual Information (NPMI) (Newman et al., 2010) measures topic coherence internally using a sliding window to count word co-occurrence patterns; Topic Coherence ($C_{V}$) (Röder et al., 2015) is a variant of NPMI that uses the one-set segmentation to count word co-occurrences and the cosine similarity as the similarity measure.

3 Results & Analysis

Directly clustering high-quality sentence embeddings can generate good topics. From Table 2, it can be observed that SBERT and SimCSE-based clustering models achieve the best averaged topic coherence among the three datasets while maintaining remarkable topic diversities. Conversely, clustering RoBERTa achieves similar or worse results than contextualized NTMs. The results suggest that contextualized embeddings are essential to get high-quality topics.

Topic words weighting method is vital. We can see in Figure 2 that inappropriate word selecting methods ($\mathbf{TFIDF\times TF_{i}}$ and $\mathbf{TF_{i}}$) lead to worse topic coherence than the contextualized NTMs (i.e., CombinedTM and ZeroShotTM), and even the BoW-based ProdLDA. Moreover, from Table 2, BERT${}_{\texttt{base}}$+KM adopt TF to obtain top words for each cluster, which ignores that the words may also be prevalent in other clusters, thus having poor topic diversities. It is also worthy to note that although BERT${}_{\texttt{base}}$+UMAP+HDBSCAN (i.e. BERTopic) reaches the highest topic diversity on 20Newsgroups, it cannot produce the specified topic numbers. Thus its performance may be boosted because of the reduced topic numbers. Moreover, our proposed methods, i.e. BERT${}_{\texttt{base}}$ and BERT${}_{\texttt{base}}$+UMAP outperforms BERTopic in most metrics, especially on topic coherence. This suggests that c-TF-IDF tends to discover incoherent words from each cluster to maintain a high topic uniqueness. Instead, our proposed method, $\mathbf{TFIDF\times IDF_{i}}$, considers the locally important words and globally infrequent words at the same time. We provide more comparison of the word selecting methods in Section 4.4.

Clustering-based topic models are robust to various lengths of documents. From Table 2 and Figure 2, we find that clustering-based models with high-quality embeddings (SBERT and SimCSE) consistently perform better than conventional LDA and NTMs, especially on the short text dataset M10, even with different word selecting methods.

4 Ablation Studies

We further investigate the impact of the topic word selecting methods, different embedding dimensionalities, as well as the topic numbers.

Topic word selecting methods. Table 3 shows the comparison between different word weighting methods. $\mathbf{TFIDF\times IDF_{i}}$ achieves significantly better results among all methods. This indicates that $\mathbf{TFIDF}$ marks out the important words to each document in the entire corpus, while $\mathbf{IDF_{i}}$ penalizes the common words in multiple clusters. Conversely, the other three methods ignore that frequent words in a cluster may also be prevalent in other clusters, hence selecting such words leading to low topic diversities. A further analysis in Appendix B also supports the observation.

Embedding dimensionality reduction. We apply UMAP to reduce the dimensionality of the sentence embeddings before clustering. As shown in Figure 3, the embeddings dimensionality negligibly affects topic quality for all word selecting methods. However, reducing to a lower dimensionality decreases the computational runtime as shown in Table 4. We compare the model runtime between the contextualized NTM CombinedTM and clustering-based models. We reduce the dimensionality of the sentence embeddings to 50 using UMAP. All models run on NVIDIA T4 GPU.

Topic numbers $K$. We investigate the impact of the different number of topics $K$ on the performance of the models. Figure 2 plots the trends of TU and $C_{V}$ on three datasets. We observe that the TU of clustering-based topic models, especially the models using $\mathbf{TFIDF\times IDF_{i}}$, decrease slowly compared to others when $K$ increases. The similar trend can be observed for topic coherence, while the $C_{V}$ of LDA and NTMs either fluctuates significantly or stays at a low level.

Conclusion

We conduct a thorough empirical study to show that a clustering-based method can generate commendable topics as long as high-quality contextualized sentence embeddings are used, together with an appropriate topic word selecting strategy. Compared to neural topic models, clustering-based models are more simple, efficient and robust to various document lengths and topic numbers, which can be applied in some situations as an alternative.

Acknowledgement

This work is fully funded by TPG Telecom. We thank anonymous reviewers for their valuable comments. We also thank Yunqiu Xu for discussions.

References

Appendix A Configuration Details

We implement LDA and NTMs based on OCTIS (Terragni et al., 2021) https://github.com/MIND-Lab/OCTIS and use their default settings. Specifically, ProdLDA, CombinedTM, and ZeroShotTM share the same configurations, i.e. one hidden layer with 100 neurons, ADAM optimizer and Momentum as 0.99; we randomly dropout 20% hidden units; we run 100 epochs of each model, and the batch size is 64. For BERT+KM, we follow Sia et al. (2020) by reducing embedding dimension to 50 using Principal Component Analysis (PCA) and adopting TF to select words. For BERT+UMAP+HDBSCAN, we follow BERTopic Grootendorst (2020) and allows it to reduce the topic numbers. For our methods, we implement clustering-based experiments based on BERTopic (Grootendorst, 2020) https://github.com/MaartenGr/BERTopic. We reduce embedding dimension to 5 using UMAP. We use BERT, RoBERTa, and SBERT embeddings provided by HuggingFace https://huggingface.co/models, and SimCSE embeddings provided from its official Github https://github.com/princeton-nlp/SimCSE.

Appendix B Comparison of Topic Words

We run Sup-SimCSE(RoBERTa${}_{\texttt{base}}$)+UMAP on 20Newsgroup and show the differences of topic diversities produced by distinct word selecting methods in Table 5. It is clear that $\mathbf{TFIDF_{i}}$ and $\mathbf{TF_{i}}$ tend to choose common words across multiple topics.