We consider an automated assistant that has been deployed in a large multinational organization to answer employee FAQs. The system is based on an LSTM classifier that has been trained on a corpus of questions and answers carefully prepared by a small team of domain experts. We find that linguistic training bias creeps into the manually created training data due to specific phrases being used, with little or no variation, which biases the deep-learning classifier towards incorrect features. Further, often the FAQs as envisaged by the trainers are in fact incomplete, and transferring linguistic variations across questionanswer pairs can uncover new question classes for which answers are missing. In this paper we demonstrate that an LSTM-based variational auto-encoder (VAE) can be used to automatically generate linguistically novel questions, which, (a) corrects classifier bias when added to the training set, (b) uncovers incompleteness in the set of answers and (c) improves the accuracy and generalization abilities of the base LSTM classifier, enabling it to learn from smaller training sets.
The recent successes of deep learning techniques for NLP have seemingly obviated the need for careful crafting of features based on linguistic properties. Deep models purportedly can learn the features required for a task directly from data, provided it comes in sucient volume and variety, typically obtained in ‘natural’ settings such as the web. However, in practical applications as we shall describe, training data is far more limited and is often created manually in a curated manner specifically for the task at hand. We show that linguistic training bias often creeps into such training data and degrades the performance of deep models. Further, manual curation can often result in incomplete task specification due to insucient linguistic variation in the training data.
We have created and deployed a chatbot for the use of employees in our
large organization, which answers human resource (HR) policy related questions
in natural language. A deep-learning model in the form of an Long short-term
memory (LSTM) [
We observed that the chosen deep-learning technique performed better than traditionally known approaches on a given test set (also created by the same HR ocers). However, the deep model also sometimes classified a user query using clearly irrelevant features (i.e., words). For example, the question “when my sick leave gets credited ? ” was classified into a category related to ‘Adoption Leave’ resulting in a completely irrelevant answer. Upon analysis we discovered that this happens mainly because the words surrounding ‘sick leave’ (e.g., “gets”) in the query occurred more often in the training data for ‘Adoption Leave’. As a result, if such words occur in users’ query, the model ignores other important words (such as ‘sick leave’) and classifies the query into incorrect class, based on such words. We refer to this phenomenon as linguistic training bias, which creeps in merely because templates used for di↵erent classes are not exhaustive. Such examples led us to fear that such a system would not generalize well when exposed to real users.
More generally, relying on human curation often results in such linguistic training biases creeping into the training data, since every individual has a specific style of writing natural language and uses some words in specific context only. Deep models end up learning these biases, instead of the core concept words of the target classes.
In order to correct these biases we automatically generate meaningful
sentences using a generative model, and then use them for training the
classification model after suitable human annotation. We use a Variational Autoencoder
(VAE) [
Further, the HR ocers created the classes and corresponding questions based on their experience of what are the frequently asked questions by various employees of the organization. However, when a chatbot is available, users tend to ask extra questions not asked otherwise. It is therefore imperative to have broader coverage of such classes. If we could show novel classes of questions to the HR ocers, generated automatically, it becomes easier for them to accept or reject the proposed classes rather than having to imagine all such possibilities. As we demonstrate in Section 6.2 our approach generated many new classes that were accepted by the HR ocers for deployment.
More specifically, the use of a deep generative model to augment training sentences resulted in the following important benefits, which form the key contributions of this paper: 1. Augmenting training data with automatically generated sentences is able to correct over-fitting due to linguistic training bias. To show this we present results on ‘concept words’, indicating potentially better generalization. 2. The newly generated sentences sometimes belonged to completely new classes not present in the original training data: 33 new classes were found in practice. 3. Augmenting training data with automatically generated sentences results in an improved accuracy (2%) of the deep-learning classifier.
We also present an improved approach for training VAEs: weighted cost annealing, based on our experience of training LSTM-VAE on a real-life dataset. 2
VAEs have found widespread usage in image generation [
As shown in [
We assume that a dataset of frequently asked questions for building a chatbot comprises of sets of semantically similar questions Qi = {q1, ..., qni } and their corresponding answer ai. A set of such questions Qi and corresponding answer ai are collectively referred to as a query set si = {Qi, ai}. Questions of a query set si are represented as Qi = Q(si). We assume that the dataset D comprises of many such query sets, i.e., D = {s1, ..., sm}. In our chatbot implementation, given a user’s query q, the objective is to select the corresponding query set s via a multi-class classification model, such that corresponding answer a could be shown.
Given all the questions in the training data D, Q = [ Q(si), 8 si 2 D, we intend to generate new questions Q0 using LSTM-VAE. Some of the questions in Q0 are semantically similar to one of the query sets of D, while the remaining questions do not belong to any of the existing classes. Using these newly generated queries Q0, we present the results of our study comprising of a) analysis of the new query-sets s0 and results of the review of these query sets by HR domain experts, b) comparison of classifiers trained on Q, and Qnew (= [{ Q, Q0}, 8 q 2 Qnew9 i, s.t. q 2 si) on core concept words rather than the full natural language questions, and c) also analyze how the accuracy of a classification model improves. 4
Recurrent neural networks (RNNs) are known for modeling sequential
dependencies in the input, for e.g. in language modeling tasks where prediction of a
word depends on the previous words in the sentence. Generally vanilla RNNs
do not handle long term dependencies in the input due to the vanishing
gradient problem [
Autoencoders are neural networks used for learning lower dimension representations of data, referred to as encoding or latent representation (z). Autoencoders are comprised of two components: an encoder (z = fenc(x)) and a decoder (xˆ = fdec(z)), where the encoder transforms the input x to z and the decoder tries to reconstruct the x from z. Autoencoders are trained by minimizing the reconstruction error of the decoded data xˆ with respect to input x, i.e., L(x, xˆ) = kx xˆk2.
The choice of encoder and decoder networks generally depends on the input
data. A sequence autoencoder uses RNNs to encode (as well as to decode)
sequential input data, e.g., a sentence. Latent representations of data, learned by
the sequence autoencoder, can also be used for classification. [
The VAE is a generative model which unlike sequence autoencoders, is comprised of a probabilistic encoder (q (z|x), recognition model ) and a decoder (p✓ (x|z), generative model ). The posterior distribution p✓ (z|x) is known to be computationally intractable. VAE approximates p✓ (z|x) using the encoder q (z|x), which is assumed to be Gaussian and is parameterized by = {µ, }, and the encoder learns to predict . As a result, it becomes possible to draw samples from this distribution.
In order to decode a sample z drawn from q (z|x), to the input x, the reconstruction loss also needs to be minimized. The reconstruction loss is represented by Eq (z|x)(log p✓ (x|z)). If the VAE were trained similar to a sequence autoencoder, i.e., with reconstruction loss only, it would not allow enough variance in q (x|z), as a result the VAE would approximately map input x into a latent representation deterministically. This gets avoided by minimizing the reconstruction loss together with the KL-divergence between q (z|x) and the prior p✓ (z).
L(, ✓, x) =
Here, p✓ (z) is assumed to be multi-variate Gaussian N (0, I). [
Sampling of z from q (z|x) is a non-continous operation, therefore it is not
possible to train the encoder via back-propagation. The Re-parametrization trick
introduced in [
N 1 X log(p(wi+1 | hi, wi))
N i=1
Generally performance of the RNNLM is measured using perplexity (lower is
better), Perplexity = expLCE LM . We have used RNNLM with LSTM units as
described in [
Classification can be considered as a two step process with the first step requiring a representation of the data. The second step involves using this representation for classification. Data can be represented using a bag of words approach, which ignores the word order information; or using hand-crafted features, which fail to generalize to multiple datasets / tasks. We learn the task-specific sentence representation using RNNs with LSTM units by representing the variable length sentence in a fixed length vector representation h, obtained after passing the sentence through the RNN layer. We then apply softmax over the ane transformation of h, i.e., p(c | h) = softmax(Wsh + bs). To learn the weights of the above model, we minimize the categorical cross entropy loss, i.e., LCE = Pim=1 y · log(p(ci | h)), where ci is one of the m class. Here, y is 1 only for the target class and zero otherwise. 5
Similar to [
After sampling and via the re-parameterization trick (explained later in this section), the sampled encoding z passes through a feed-forward layer to obtain hd0 , which is the start state of the decoder RNN. We pass the encoding z as input to the LSTM layer at every time-step. The vector representation of the word with the highest probability, predicted at time t, is also passed as input at next time-step (t + 1), as shown in Figure 1.
Training the LSTM-VAE It is not straightforward to train the LSTM-VAE using the loss function given in Eq. 1. As described in Section 4.2, if the KLdivergence loss term is not used, the model will converge to a state that gives fixed latent representations. Similarly, if the reconstruction loss is high it would lead to meaningless sentences. If equal weightage is given to both loss terms the Source Sent: where can i apply earned leave? Generated Sentences where can i apply earned leave ? where can can check lwp guidelines ? where can i apply sick leave ? where can i apply earned leave ? where can i apply earned leave ? where can i find lwp guidelines ? where can i apply earned leave ? where can can apply casual leave ? how can i apply earned leave ? where can i find lwp leave ? model learns to encode the input sequence into a desired latent representation too well, i.e., KL divergence (KLD) between approximate posterior q (z|x) and prior p✓ (z) reaches zero. As a result, the decoder network reduces to merely a language model (RNNLM) and it does not generate novel samples. The network should be trained such that both of the loss terms gradually decrease towards convergence.
This problem was also reported by [
– Weighted cost annealing: We have used an improved version of the KL cost
annealing presented in [
Inference: Sentence Generation To generate sentences similar to input
sentences, we have used the recognition model to get the parameters of the
distribution corresponding to a sentence (µ, ). If we sample z from this distribution, the
sampling step will become non-continuous and we will not be able to learn the
parameters of the neural network using back propagation. The re-parameterization
trick [
z = µ + ✏ · , where ✏ ⇠ N (0, I) (4) 5.2
It is dicult to use the sentences generated by the LSTM-VAE in a downstream
task (classification in our case) without preprocessing because of the word
repetition problem in the generated samples observed by various researchers [
Base Classifier (M1): We use the single layer recurrent neural network with LSTM units for classification trained on the original data. We use this as a baseline for the classification task.
Base Classifier augmented with generated sentences (M2): To obtain the label for the novel sentences generated by VAE, we use M1 and choose the top N sentences, based on the entropy of the softmax distribution, as candidates for augmenting the training data. We manually verify the label and correct the label if it is incorrectly classified by M1; we also remove the sentences that clearly correspond to new classes. We augment the training data with this new dataset and train another classification model (M2), as shown in Figure 2.
Base Classifier augmented with generated sentences based on word
embedding(M3): Here, we try to augment the training set with additional
sentences that are generated by replacing some words in each training sentence. All
the words in a sentence except english stopwords and leave types {sick, casual
etc.} (to avoid change of class label) are selected as the candidate words for
replacement. We replace a candidate word with a word corresponding to the
nearest neighbour of the candidate word’s embedding. The nearest neighbour is
chosen based on the cosine similarity between two word embeddings obtained
from glove [
6.1
Dataset Description : We use a dataset of Leave policy questions, created by HR ocers 1, for chatbot creation, see Table 2 for more details. This dataset was divided in three parts (Training, Validation and Test) in 60:20:20 ratio. Training and Validation datasets are used for training the LSTM-VAE model, here we select the best model based on loss on validation data, i.e., L(, ✓, x), see Eq. 3. New sentences are generated using only training data as input. The test data is used for reporting the performance metrics of the classification model, and is never exposed to LSTM-VAE model. Another test data comprising only the core concept words was also used for testing (see Section 6.3).
Characteristic N c l
V Value
Training Details: Word embeddings for tokens (delimited by space) are
initialized randomly and learned during the training. We use Adam [
Parameter Dim. of word embed.
Dimension of z (30) weighted cost annealing When preparing the questions to train a chatbot as described in Section 3, it is hard to visualize all possible questions (i.e., query-sets) that users can ask when it is made live. Using LSTM-VAE we were able to generate many new query sets, which were accepted by the core HR Team. In Figure 2 we present the entire workflow followed to generate the novel queries. Out of the 175,000 queries generated using LSTM-VAE, we removed the queries already present in the training data, as well as those which have same word repeating more than once consecutively. After this stage we obtained about 5,700 sentences. These sentences were then tested using a LM (as described in Section 4.3), and we picked-up only top 1500 sentences based on the likelihood. Many of these sentences were found to be grammatically correct, while only some of them were semantically inconsistent. In this process it was discovered that 434 sentences did not belong to any of the existing classes. These sentences were given to the HR team for review, and they selected 120 sentences belonging to 33 new classes. We have made this chatbot live in our company and these 33 classes are in practical use (see Key Contribution 2 ).
Further, in Table 4, we show sample novel sentences generated by LSTMVAE belonging to 9 di↵erent classes (separated by horizontal line in the table). For every novel sentence, we also show the most similar sentence taken from training data, identified using Jaccard Similarity. Next, our HR team tested the chatbot built using models M1 and M2 (see Section 5.3) by using only core concept words as inputs, rather than complete sentences: see Table 5. To build the model M2, we added sentences generated by the process flow, shown in Figure 2. Here, out of 1500 sentences chosen by the LM 1066 were manually found to be belonging to existing classes, and we ran the model M1 Sentences from training data Are business associates eligible to take al? Are business associates eligible for al? Where can i find the policy on casual leave? Where can i read more about casual leave ? Where do i get more info on casual leaves? Where i can find cl guidelines ? How many flexi leaves will i get ? What is flexi leave and how can i avail them ? Can i apply ml if i am on lwp ? How can i go about applying ml if on lwp ? Is it possible for me to apply ml when on lwp ? What happens to my unutilized sick leave ? In event of separation can i encash my sick leave ? Concept Word Query Adoption leave business associates Casual leave policy path Flexi Leave Entitlement Maternity leave entitlement while on LWP
Sick leave resign during separation ? on these sentences. We chose top-250 correctly classified sentences based on low entropy and all the 130 wrongly classified sentences, and combined these with the training data to train the model M2. These 380 sentences belong to only 74 out of 117 classes. As shown in Table 6, a classifier built using only the training dataset tends to have linguistic training bias. This bias gets reduced when the generated sentences are also used (see Key Contribution 1 ). This di↵erence of 9% can be attributed to new sentences generated via perturbation of surrounding words and high quality sentences generated by LSTM-VAE. 6.4
We now analyze the impact of addition of this training data into the model on hold-out set. Here, a small gain of about 2% was observed in classification accuracy, as shown in Table 6 (see Key Contribution 3 ). While this gain is small, it is important that the classifier uses the right words; especially since the test set, being a subset of curated questions, also has linguistic training bias. It was observed that the model M1 often classified the query using non-relevant words as shown in Table 6. Even if such a classifier had worked better on the test data, we believe it would perform worse on actual user queries; of course this can only be confirmed by a controlled A/B test on real users. 6.5
During user testing of the FAQ-bot, it was observed that often model’s decision to classify a user query in certain class is based on non-concept words. It was suspected that this phenomenon occurs due to linguistic training bias, as explained Dataset
M1
M 2
M 3 Concept Word Testing 55.00% 64.28% 62.14% Test Set in Section 1. New sentences generated by LSTM-VAE, which were wrongly classified by M1, are likely to be of this type, and are therefore more value adding for the model. For example, many sentences get generated which contain the same surrounding words (e.g., ‘gets’) but belonging to di↵erent classes, e.g., “ When my casual leaves gets credited ? ” and “When my maternity leaves gets credited ? ”. When such sentences are added to the training set, it indirectly forces the model to learn to distinguish the classes based on some other words than such non-concept words. Perhaps this is one of the reasons why merely with an addition of 130 sentences (which is less than 10% of the original training data) accuracy gain of almost 10% was observed during concept-word based testing of the model as shown in Table 6. It would otherwise have required many more new sentences to get the same degree of gain in classification accuracy. 7
Based on our experience of building and deploying a FAQ-bot in practice, we have noted that linguistic training bias leads to over-fitting of deep-learning model. Such bias is more likely to occur when the training data is created by domain experts, who tend to use templates when writing natural language sentences. We have shown that such bias can be fixed to certain extent by generating novel sentences using a generative model. We also described an approach for such a generative model, which uses LSTM-VAE followed by sentence selection using a LM. We presented weighted cost annealing, an improved approach for training the LSTM-VAE, and showed that accuracy of a classification model can be increased significantly by using generated sentences. Most importantly, we have shown that our approach was able to generate newer classes of questions for our FAQ-chatbot, not present in the original training data, which were reviewed and accepted by the domain experts for deployment.
– Weighted cost annealing: As mentioned in Section 5.1 the training of LSTM-VAE by minimizing the loss according to Equation 5 causes the KLdivergence loss to increase for few time-steps initially and then drop to zero as shown in Figure 3. This makes the decoder to behave like an RNNLM. To overcome the above issue we use weighted cost annealing i.e. we increase the weight of KL-divergence loss lineraly after every e epochs and simultaneously reduce the weight of the reconstruction loss. Due to this, even though the KL-divergence loss increases initially for few time-steps, it starts decreasing over the time-steps but remains non zero as shown in Figure 3. x) = KL(q (z|x)kp✓ (z)) – As described in Section 5.1 the input to the LSTM-VAE decoder, for predicting the word at time t + 1, is the output of the decoder at time t rather than the actual word from the input sentence. In our case we found that this helps the LSTM-VAE in generating more semantically and syntactically correct sentences as compared to those generated by inputting the actual word to the decoder. This can be observed from the Table 7.