Multi-View Domain Adapted Sentence Embeddings for Low-Resource Unsupervised Duplicate Question Detection

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1630–1641, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1630 Multi-View Domain Adapted Sentence Embeddings for Low-Resource Unsupervised Duplicate Question Detection Nina Poerner1,2 and Hinrich Sch¨utze1 1Center for Information and Language Processing, LMU Munich, Germany 2Corporate Technology Machine Intelligence (MIC-DE), Siemens AG Munich, Germany poerner@cis.uni-muenchen.de | inquiries@cislmu.org Abstract We address the problem of Duplicate Question Detection (DQD) in low-resource domain- specific Community Question Answering fo- rums. Our multi-view framework MV-DASE combines an ensemble of sentence encoders via Generalized Canonical Correlation Anal- ysis, using unlabeled data only. In our ex- periments, the ensemble includes generic and domain-specific averaged word embeddings, domain-finetuned BERT and the Universal Sentence Encoder. We evaluate MV-DASE on the CQADupStack corpus and on addi- tional low-resource Stack Exchange forums. Combining the strengths of different encoders, we significantly outperform BM25, all single- view systems as well as a recent supervised domain-adversarial DQD method. 1 Introduction Duplicate Question Detection is the task of finding questions in a database that are equivalent to an incoming query. Many Community Question An- swering (CQA) forums leave this task to the col- lective memory of their users. This results in un- necessary manual work for community members as well as delayed access to answers (Hoogeveen et al., 2015). Automatic DQD is often approached as a super- vised problem with community-generated training labels. However, smaller CQA forums may suf- fer from label sparsity: On Stack Exchange, 50% of forums have fewer than 160 user-labeled dupli- cates, and 25% have fewer than 50 (see Figure 1).1 To overcome this problem, two avenues have been explored: The first is supervised domain- adversarial training on a label-rich source fo- rum (Shah et al., 2018), which works best when 1archive.org/details/stackexchange [data dump: 2018-12-20] 100 101 102 103 104 105 106 #D #Q Figure 1: Distribution (log-scale box plot) of number of questions (#Q) and number of labeled duplicates (#D) on Stack Exchange. N = 165 forums. source and target domains are related. The sec- ond is unsupervised DQD via representation learn- ing (Charlet and Damnati, 2017; Lau and Baldwin, 2016), which requires only unlabeled questions. In this paper, we take the unsupervised avenue. A major challenge in the context of domain- specific CQA forums is that language usage may differ from the “generic” domains of existing rep- resentations. To illustrate this point, compare the following Nearest Neighbor lists of the word “tree”, based either on generic GloVe embeddings (Pennington et al., 2014) or on FastText embed- dings (Bojanowski et al., 2017) that were trained on specific CQA forums: generic (GloVe): trees, branches, leaf chess: searches, prune, modify outdoors: trees, trunk, trunks gis: strtree, rtree, btree wordpress: trees, hierachy, hierarchial gaming: trees, treehouse, skills Charlet and Damnati (2017) and Lau and Bald- win (2016) report that representations trained on in-domain data perform better on unsupervised DQD than generic representations. But in a low-resource setting, …

Content truncated to 500 words