CIKM 2019
In this paper, we propose a novel method for a sentence-level answer-selection task that is one of the fundamental problems in natural language processing. First, we explore the effect of additional information by adopting a pretrained language model to compute the vector representation of the input text and by applying transfer learning from a large-scale corpus. Second, we enhance the compare-aggregate model by proposing a novel latent clustering method to compute additional information within the target corpus and by changing the objective function from listwise to pointwise. To evaluate the performance of the proposed approaches, experiments are performed with the WikiQA and TRECQA datasets. The empirical results demonstrate the superiority of our proposed approach, which achieve state-of-the-art performance on both datasets.
CTAC 2024 Valencia - Sven Zoelle - Most Crucial Invest to Digitalisation_slid...
[poster] A Compare-Aggregate Model with Latent Clustering for Answer Selection
1. A Compare-Aggregate Model with Latent Clustering
for Answer Selection
Seunghyun Yoon1†, Franck Dernoncourt2, Doo Soon Kim2, Trung Bui2 and Kyomin Jung1
1Seoul National University, Seoul, Korea 2Adobe Research, San Jose, CA, USA
mysmilesh@snu.ac.kr
Research Problem
• We propose a novel method for a sentence-level answer-selection task.
• We explore the effect of additional information by adopting a pretrained lan-
guage model to compute the vector representation of the input text and by
applying transfer learning from a large-scale corpus.
• We enhance the compare-aggregate model by proposing a novel latent clus-
tering method and by changing the objective function from listwise to point-
wise.
Methodology
Latent memory
𝑴𝑴1 … 𝑴𝑴𝑛𝑛
Sentence
Representation
Compute
Similarity
𝑴𝑴LC
𝐴𝐴
= � 𝛼𝛼𝑘𝑘
𝐴𝐴
∙ 𝑴𝑴𝑘𝑘
𝑘𝑘
𝒑𝒑1:𝑛𝑛
𝐴𝐴
= (𝒔𝒔𝐴𝐴
)⊺
𝑾𝑾𝑾𝑾
𝒔𝒔 𝐴𝐴
=
1
𝑚𝑚
� 𝒂𝒂�𝑖𝑖
𝑖𝑖
𝑞𝑞1 𝑞𝑞2 𝑞𝑞𝑛𝑛
Context
Representation
𝒍𝒍1
𝑄𝑄
𝒍𝒍1
𝑄𝑄
𝒍𝒍2
𝑄𝑄
𝒍𝒍2
𝑄𝑄
𝒍𝒍𝑛𝑛
𝑄𝑄
𝒍𝒍𝑛𝑛
𝑄𝑄
Comparison
+
𝒒𝒒�1 𝒒𝒒�2 𝒒𝒒�𝑛𝑛
Language
Model
Attention
𝒉𝒉1
𝑄𝑄
𝒂𝒂�1
⊙
𝒉𝒉2
𝑄𝑄
𝒂𝒂�2
⊙
𝒉𝒉 𝑚𝑚
𝑄𝑄
𝒂𝒂� 𝑚𝑚
⊙
𝒄𝒄1
𝑄𝑄
CNN
Latent
Clustering
Aggregation
𝒍𝒍1
𝐴𝐴
𝒍𝒍1
𝐴𝐴
𝒍𝒍2
𝐴𝐴
𝒍𝒍2
𝐴𝐴
𝒍𝒍 𝑚𝑚
𝐴𝐴
𝒍𝒍 𝑚𝑚
𝐴𝐴
+
𝒂𝒂�1 𝒂𝒂�2 𝒂𝒂� 𝑚𝑚
𝒉𝒉1
𝐴𝐴
𝒒𝒒�1
⊙
𝒉𝒉2
𝐴𝐴
𝒒𝒒�𝟏𝟏
⊙
𝒉𝒉𝑛𝑛
𝐴𝐴
𝒒𝒒�𝑛𝑛
⊙
CNN
𝑎𝑎1 𝑎𝑎2 𝑎𝑎 𝑚𝑚
𝒄𝒄2
𝑄𝑄
𝒄𝒄 𝑚𝑚
𝑄𝑄
𝒄𝒄1
𝐴𝐴 𝒄𝒄2
𝐴𝐴 𝒄𝒄𝑛𝑛
𝐴𝐴
…
…
…
…
…
…
…
…
…
…
softmax
Latent-cluster
Information
𝒔𝒔𝐴𝐴
𝑴𝑴LC
𝐴𝐴
k-max-pool
𝒑𝒑�1:𝑘𝑘
𝐴𝐴
𝛼𝛼1:𝑘𝑘
𝐴𝐴
Figure 1: The architecture of the model. The dotted box on the right shows the process through which the latent-cluster infor-
mation is computed and added to the answer. This process is also performed in the question part but is omitted in the figure.
The latent memory is shared in both processes.
• Pretrained Language Model (LM): We select an ELMo language model and
replace the previousword embedding layer with the ELMo model.
• Latent Clustering (LC): We assume that extracting the LC information of the
text and using it as auxiliary information will help the neural network model
analyze the corpus.
p1:n = s WM1:n,
p1:k = k-max-pool(p1:n),
α1:k = softmax(p1:k),
MLC = kαkMk.
(1)
MQ
LC = f(( iqi)/n), qi ⊂ Q1:n,
MA
LC = f(( iai)/m), ai ⊂ A1:m,
CQ
new = [CQ
; MQ
LC], CA
new = [CA
; MA
LC].
(2)
• Transfer Learning (TL): To observe the efficacy in a large dataset, we apply
transfer learning using the question-answering NLI (QNLI) corpus.
• Point-wise Learning: Previous research adopts a listwise learning approach
using KL-divergence score. In contrast, we pair each answer candidate to the
question and compute the cross-entropy loss to train the model.
scorei = model(Q, Ai),
S = softmax([score1, ..., scorei]),
loss = N
n=1KL(Sn||yn).
(3)
loss = − N
n=1 yn log (scoren). (4)
Datasets
• WikiQA (Yang et al. 2015) is an answer selection QA dataset constructed from
real queries of Bing and Wikipedia
• TREC-QA (Wang et al. 2007) is an answer selection QA dataset created from
the TREC Question-Answering tracks
• QNLI (Wang et al. 2018) is a modified version of the SQuAD dataset that
allows for sentence selection QA
Dataset
Listwise pairs Pointwise pairs Answer candidates
/ question
train dev test train dev test
WikiQA 873 126 143 8.6k 1.1k 2.3k 9.9
TREC-QA 1.2k 65 68 53k 1.1k 1.4k 43.5
QNLI 86k 10k - 428k 169k - 5.0
Table 1: Properties of the dataset.
Experimental Results
We regard all tasks as relevant answer selections for the given questions. Follow-
ing the previous study, we report the model performance as the mean average
precision (MAP) and the mean reciprocal rank (MRR).
# Clusters
Listwise Learning Pointwise Learning
MAP MRR MAP MRR
1 0.822 0.819 0.842 0.841
4 0.839 0.840 0.846 0.845
8 0.841 0.842 0.846 0.846
16 0.840 0.842 0.847 0.846
Table 2: Model (Comp-Clip +LM +LC) performance on the QNLI corpus with a variant number of clusters.
Conclusions
• Achieve state-of-the-art performance on the WikiQA and TREC-QA
• Show that leveraging a large amount of data is crucial for capturing the con-
textual representation of input text
• Show that the proposed latent clustering method with a pointwise objective
function significantly improves model performance
Model
WikiQA TREC-QA
MAP MRR MAP MRR
dev test dev test dev test dev test
Compare-Aggregate (Wang et al. 2017) [1] 0.743* 0.699* 0.754* 0.708* - - - -
Comp-Clip (Bian et al. 2017) [2] 0.732* 0.718* 0.738* 0.732* - 0.821 - 0.899
IWAN (Shen et al. 2017) [3] 0.738* 0.692* 0.749* 0.705* - 0.822 - 0.899
IWAN + sCARNN (Tran et al. 2018) [4] 0.719* 0.716* 0.729* 0.722* - 0.829 - 0.875
MCAN (Tay et al. 2018) [5] - - - - - 0.838 - 0.904
Question Classification (Madabushi et al. 2018) [6] - - - - - 0.865 - 0.904
Listwise Learning to Rank
Comp-Clip (our implementation) 0.756 0.708 0.766 0.725 0.750 0.744 0.805 0.791
Comp-Clip (our implementation) + LM 0.783 0.748 0.791 0.768 0.825 0.823 0.870 0.868
Comp-Clip (our implementation) + LM + LC 0.787 0.759 0.793 0.772 0.841 0.832 0.877 0.880
Comp-Clip (our implementation) + LM + LC +TL 0.822 0.830 0.836 0.841 0.866 0.848 0.911 0.902
Pointwise Learning to Rank
Comp-Clip (our implementation) 0.776 0.714 0.784 0.732 0.866 0.835 0.933 0.877
Comp-Clip (our implementation) + LM 0.785 0.746 0.789 0.762 0.872 0.850 0.930 0.898
Comp-Clip (our implementation) + LM + LC 0.782 0.764 0.785 0.784 0.879 0.868 0.942 0.928
Comp-Clip (our implementation) + LM + LC +TL 0.842 0.834 0.845 0.848 0.913 0.875 0.977 0.940
Table 3: Model performance (the top 3 scores are marked in bold for each task). We evaluate model [1-4] on the WikiQA corpus using author’s implementation (marked by *). For TREC-QA case, we present reported results in the original papers.
†
Work conducted while the author was an intern at Adobe Research.