Evaluating Coherence in Dialogue Systems using Entailment

TL;DR

This paper proposes new automatic, interpretable metrics based on entailment and sentence representations to evaluate dialogue coherence, aiming to replace costly human judgments and improve scalability and consistency.

Contribution

It introduces novel entailment-based metrics for assessing dialogue coherence, providing a scalable and unbiased alternative to human evaluation.

Findings

Metrics correlate well with human judgments

Enables large-scale evaluation of dialogue systems

Reduces evaluation bias and cost

Abstract

Evaluating open-domain dialogue systems is difficult due to the diversity of possible correct answers. Automatic metrics such as BLEU correlate weakly with human annotations, resulting in a significant bias across different models and datasets. Some researchers resort to human judgment experimentation for assessing response quality, which is expensive, time consuming, and not scalable. Moreover, judges tend to evaluate a small number of dialogues, meaning that minor differences in evaluation configuration may lead to dissimilar results. In this paper, we present interpretable metrics for evaluating topic coherence by making use of distributed sentence representations. Furthermore, we introduce calculable approximations of human judgment based on conversational coherence by adopting state-of-the-art entailment techniques. Results show that our metrics can be used as a surrogate for human judgment, making it easy to evaluate dialogue systems on large-scale datasets and allowing an unbiased estimate for the quality of the responses.

Tables

Table 1. Table 1: Distribution of labels in the InferConvAI corpus.

	Train	Dev	Test
#entailment	218.2K	12.2K	1.4K
#neutral	579.5K	28.0K	3.1K
#contradiction	261.9K	9.8K	1.1K
Total	1.1M	50.2K	5.6K

Table 2. Table 2: Accuracy of inference models on InferConvAI.

Method	Reddit	OpenSubtitles
ESIM + ELMo	0.526	0.455
BERT	0.553	0.498

Table 3. Table 3: The Pearson Correlation between different metrics and human judgments with p 𝑝 p -value < 0.001 absent 0.001 <0.001 . The semantic similarity (SS) metric is measured with respect to the most recent utterance H − 1 subscript 𝐻 1 H_{-1} and the most recent two utterances H − 2 subscript 𝐻 2 H_{-2} in the conversation history. We adopt different embedding algorithms to compute the word vectors: ELMo Peters et al. ( 2018 ) , BERT Devlin et al. ( 2018 ) , word2vec Mikolov et al. ( 2013 ) and Universal Sentence Encoder (USE) Cer et al. ( 2018 ) .

Method	Pearson
	Reddit	OpenSubtitles
SS${\left(H_{-2}\right)}_{\mathrm{BERT}}$	-0.204	-0.290
SS${\left(H_{-2}\right)}_{\mathrm{ELMo}}$	-0.146	-0.365
SS${\left(H_{-2}\right)}_{\mathrm{USE}}$	-0.248	-0.314
SS${\left(H_{-1}\right)}_{\mathrm{BERT}}$	-0.214	-0.337
SS${\left(H_{-1}\right)}_{\mathrm{ELMo}}$	-0.178	-0.404
SS${\left(H_{-1}\right)}_{\mathrm{USE}}$	-0.287	-0.320
ABERT	0.135	0.131
AELMo	0.085	0.162
Aword2vec	0.037	0.196
GBERT	0.208	0.132
GELMo	0.037	0.072
Gword2vec	-0.033	0.015
EBERT	0.162	0.144
EELMo	0.035	0.116
Eword2vec	-0.065	0.118

Full text

1 Introduction

Recently, we have witnessed a big success in the capability of computers to seemingly understand natural language text and to generate plausible responses to conversations Serban et al. (2016); Xing et al. (2017); Sordoni et al. (2015); Li et al. (2016); Serban et al. (2017); Devlin et al. (2018); Radford et al. (2018). A challenging task of building dialogue systems lies in evaluating the quality of their responses. Typically, evaluating goal-oriented dialogue systems is done via human-generated judgment like a task completion test or user satisfaction score Walker et al. (1997); Möller et al. (2006). However, the task of evaluating open-ended dialogue systems is not well defined as there is no clear explicit goal for conversations. Indeed, dialog systems are ultimately created to satisfy the user’s need which can be associated with how entertaining and engaging the conversation was. It is unclear how to define a metric that can account comprehensibly for the semantic meaning of the responses. Moreover, grasping the underlying meaning of text has always been fraught with difficulties, which are essentially attributed to the complexities and ambiguities in natural language. Generally, a good dialogue can be described as an exchange of information that sustain coherence through a train of thoughts and a flow of topics. Therefore, a plausible way to evaluate open-ended dialogue systems is to measure the consistency of the responses. For example, a neural dialogue system can respond to the utterance “Do you like animals?” by “Yes, I have three cats”, thereafter replies to “How many cats do you have” by “I don’t have cats.”. Here, we can notice that the dialogue system failed to provide a coherent answer and instead generated an inconsistent response.

In this work, we characterize the consistency of dialogue systems as a natural language inference (NLI) Dagan et al. (2006) problem. In particular, NLI is focused on recognizing whether a hypothesis is inferred from a premise. In dialogue systems, we cast a generated response as the hypothesis and the conversation history as the premise, projecting thus the automatic evaluation into an NLI task. In other words, we propose directly calculable approximations of human evaluation grounded on conversational coherence and affordance by using state-of-the-art entailment techniques. For this purpose, we build a synthesized inference data from conversational corpora. The intuition behind this choice is motivated by the fact that utterances in a human conversation tend to follow a consistent and coherent flow where each utterance can be inferred from the previous interactions. We train the state-of-the-art inference models on our conversational inference data and then the learned models are used to evaluate the coherence in a given conversation. Finally, we fare our proposed evaluation method against existing automated metrics. The results highlight the capability of inference models to automatically evaluate dialogue coherence. The source code and the dataset are available at https://github.com/nouhadziri/DialogEntailment

2 Related Work

Evaluating open-ended dialogue systems has drawn the attention of several researchers in recent years. Unfortunately, word-overlapping metrics such as BLEU have been shown to correlate weakly with human evaluation, which in turn, introduces bias against certain models Liu et al. (2016). Many studies have been proposed to improve the quality of automated metrics. In particular, Lowe et al. Lowe et al. (2017) introduced an automatic evaluation system called ADEM which learns to score responses from an annotated dataset of human responses scores. However, such system is heavily biased towards the training data and struggles with generalization capabilities on unseen datasets. Further, collecting an annotated gold standard of human judgment is very expensive and thus, ADEM is less flexible and extensible. Venkatesh et al. Venkatesh et al. (2018) introduced a framework for evaluating the quality of the conversations based on topical diversity, coherence, engagement and conversational depth and showed that these metrics conform with human evaluation. However, a big part of their metrics relies on human labels, which makes the evaluation system not scalable. Recently, Welleck et al. Welleck et al. (2018) investigated the use of NLI models (e.g., ESIM Chen et al. (2016) and InferSent Conneau et al. (2017)) to measure consistency in dialogue systems. They built a Dialogue NLI dataset which consists of sentence pairs labeled as entailment, neutral, or contradiction. The utterances are derived from a two-agent persona-based dialogue dataset. To annotate the dataset, they used human annotation from Amazon Mechanical Turk. In this work, we propose a method that employs NLI approaches to detect coherence in dialogue systems. The proposed evaluation procedure does not require human labels, making progress towards scalable and autonomous evaluation systems.

3 Natural Language Inference

Reasoning about the semantic relationship between two utterances is a fundamental part of text understanding. In this setting, we consider inference about entailment as a useful testing bed for the evaluation of coherence in dialogue systems. The success of NLI models111Recent models have achieved high accuracy in Stanford NLI corpus Bowman et al. (2015) (90.1%) and GLUE Benchmark Wang et al. (2018) (86.7%) allows us to frame automated dialogue evaluation as an entailment problem. More specifically, given a conversation history $H$ and a generated response $r$, the goal is to understand whether the premise-hypothesis pair $(H,r)$ is entailing, contradictory, or neutral.

3.1 Coherence in Dialogue Systems

The essence of neural response generation models is designed by maximizing the likelihood of the target response given source utterances. Therefore, a dialogue generation task can be formulated as a next utterance prediction problem. In particular, the model predicts a response $u_{i+1}$ given a conversation history $(u_{1},...,u_{i})$. One key factor for a successful conversation is having coherence across multiple turns. A machine’s response can be considered as incoherent when it contradicts directly its previous utterances or follows an illogical reasoning throughout the whole conversation. Inconsistency can be clearly identified when it corresponds to logical discrepancy between two facts. For example, when you indicate clearly during the conversation that you have cats but when you get asked “How many cats do you have”, you answer by “I don’t have cats.”. Nevertheless, in general, inconsistency can be less explicitly recognizable as it may describe an error between what the person has said and what she/he truly believes given her/his personality and background information.

To detect dialogue incoherence, we consider two prominent models that have shown promising results in commonsense reasoning: the Enhanced Sequential Inference Model (ESIM) Chen et al. (2016) and Bidirectional Encoder Representations from Transformers (BERT) Devlin et al. (2018):

ESIM Chen et al. (2016): employs a Bi-LSTM model Graves and Schmidhuber (2005) to encode the premise and the hypothesis. Also, it explores the effectiveness of syntax for NLI by encoding syntactic parse trees of premise and hypothesis through Tree-LSTM Zhu et al. (2015). Then, the input encoding part is followed by a matrix attention layer, a local inference layer, another BiLSTM inference composition layer, and finally a pooling operation before the output layer. We further boost ESIM with by incorporating contextualized word embeddings, namely ELMo Peters et al. (2018), into the inference model.

BERT Devlin et al. (2018): exploits a multi-layer Bidirectional Transformers model Vaswani et al. (2017) to learn pre-trained universal representations of text using only a plain text corpus from Wikipedia. BERT has achieved state-of-the art results on various natural language understanding tasks and has been shown to handle strongly long-range dependencies in text. BERT can be fine-tuned to achieve several tasks by solely adding a small layer to the core model. In this work, we adopted BERT to the task of NLI.

Overall, the goal of the above models is to learn a function $G_{NLI}$ that predicts one of three categories (i.e., entailment, contradiction or neutral) given premise-hypothesis pairs.

4 Inference Corpus for Dialogues

To train the inference models, we build a synthesized dataset geared toward evaluating consistency in dialogue systems. To this end, the Persona-Chat conversational data Zhang et al. (2018) is used to form the basis of our conversational inference data. The continuity of utterances in human conversation facilitates the use of entailment in the dialogue domain. Typically, when we interact with one another, we tend to reference information from previous utterances to engage with the interlocutor. This is why we build our synthetic inference dataset upon a dialogue corpus. The Persona-Chat corpus is a crowd-sourced dataset where two people converse with each other based on a set of randomly assigned persona. To build an inference corpus, we need to find three different labels (i.e., entailment, contradiction, and neutral). For this purpose, we map an appropriate and on topic response to the entailment label. Consequently, the entailment instances are derived from the utterances in the conversations. For contradiction, grammatically-impaired sentences are constructed by randomly choosing words from the conversation. We also added randomly drawn contradictory instances from the MultiNLI corpus Williams et al. (2018) to account for meaningful inconsistencies. Finally, random utterances from other conversations or generic responses such as “I don’t know” comprise the neutral instances. Following this approach, we build a corpus of 1.1M premise-hypothesis pairs, namely InferConvAI. Table 1 summarizes the statistics of InferConvAI.

5 Experiments

In this section, we focus on the task of evaluating the next utterance given the conversation history. We used the following models to generate responses. These models were trained on the conversational datasets, using optimization, until convergence:

• •

Seq2Seq with attention mechanism Bahdanau et al. (2015): predicts the next response given the previous utterance using an encoder-decoder model.

• •

HRED Serban et al. (2016): extends the Seq2Seq model by adding a context-RNN layer that accounts for contextual information.

• •

TA-Seq2Seq Xing et al. (2017): extends the Seq2Seq model by biasing the overall distribution towards leveraging topic words in the response.

• •

THRED Dziri et al. (2018): builds upon TA-Seq2Seq model by levering topic words in the response in a multi-turn dialogue system.

The training was conducted on two datasets: OpenSubtitles Tiedemann (2012) and Reddit Dziri et al. (2018). Due to lack of resources, we randomly sampled 6M dialogues as training data from each dataset, 700K dialogues as development data, and 40K dialogues as test data. Each dialogue corresponds to three turn exchanges. To evaluate accurately the quality of the generated responses, we recruited five native English speakers. The judges annotated 150 dialogues from Reddit and 150 dialogues from OpenSubtitles. All subjects have informed consent as required from the Ethics Review Board at the University of Alberta. Due to lack of space, we will omit an exhaustive description of the human evaluation process and refer readers to Dziri et al. (2018) as we conducted the same evaluation procedure.

5.1 NLI in Dialogues

In this section, we evaluate the performance of the state-of-the-art entailment models on predicting a score for the generated utterances. In particular, the conversation history $H$ is treated as a premise, whereas the generated response $r$ acts as a hypothesis. We pick two state-of-the-art NLI models (i.e., ESIM Chen et al. (2016) and BERT Devlin et al. (2018)). These models were trained on the InferConvAI dataset. During evaluation, we use our test dialogue corpus from Reddit and OpenSubtitles, in which the majority vote of the 4-scale human rating constitutes the labels. The results are illustrated in Table 2. Both models reach reasonable performance in this setting, while BERT outperforms ESIM. Note that this experiment examines the generalization capabilities of these inference models as the test datasets are drawn from an entirely different distribution than the training corpus. Figure 1 illustrates the performance of BERT for each class with respect to the human scores. The test utterances that are predicted as entailment tend to be rated higher than other utterances, exhibiting that the entailment models correlate quite well with what humans perceive as a coherent response. Another observation is that the inference models often classify acontextual and off-topic responses as neutral and the annotators typically dislike these types of responses. This contributes to the lower scores of neutral-detected responses compared to responses predicted as contradiction.

5.2 Automated Metrics

5.2.1 Word-level Metrics

We consider as evaluation metrics baselines three textual similarity metrics Liu et al. (2016) based on word embeddings: Average (A), Greedy (G), and Extrema (E). These word-level embedding metrics have been proven to correlate with human judgment marginally better than other world-overlap metrics (e.g., BLEU, ROUGE and METEOR) Liu et al. (2016). One critical flaw of these embedding metrics is that they assume that each word is independent of the other words in the sentence. Further, the sentence is treated as a bag-of-words, disregarding words order and dependencies that are known to be substantial for understanding the semantic of a sentence. The correlation of these metrics with human judgment is showcased in Table 3. We can notice that the three metrics A, G and E correlate weakly with human judgment in both datasets, demonstrating the need for a well-designed automated metric that provides an accurate evaluation of dialogues.

5.2.2 Semantic Similarity

The Semantic Similarity (SS) metric was suggested by Dziri et al. (2018). It measures the distance between the generated response and the utterances in the conversation history. The intuition of this metric revolves around capturing good and consistent responses by showing whether the machine-generated responses maintain the topic of the conversation. In this project, we measured SS with respect to two different utterances, the conversation history $H$ and the most recent utterance $H_{-1}$. The conversation history is formed by concatenating the two most recent utterances. We report the Pearson coefficient of this metric with human judgment in Table 3. The SS metric is expected to have a negative correlation as the higher human ratings correspond to the lower semantic distance. The results demonstrate that SS metrics correlate better than word-level metrics as they make use of word interactions to represent utterances. Moreover, the Universal Sentence Encoder (USE) Cer et al. (2018) model performs better on Reddit, whereas the ELMo embeddings achieve higher correlation on OpenSubtitles. This arguably underlines that deep contextualized word representations can manage better complex characteristics of natural language (e.g., syntax and semantics). The SS metric, which requires no pre-training, reaches a Pearson correlation of -0.404 with respect to the most recent utterance on OpenSubtitles. Such correlation can be compared with a correlation of 0.436 achieved by ADEM Lowe et al. (2017) which required large amounts of training data and computation. Moreover, in order to investigate whether the results in Table 3 are in line with human evaluation, we visualized the correlation between the human ratings and SS metric as scatter plots in Figure 2.

6 Conclusion

Evaluating dialogue systems has been heavily investigated, but researchers are still on the quest for a strong and reliable metric that highly conforms with human judgment. Existing automated metrics show poor correlation with human annotations. In this paper, we present a novel paradigm for evaluating the coherence of dialogue systems by using state-of-the-art entailment techniques. We aim at building a system that does not require human annotation, which in turn, can lead to a scalable evaluation approach. While our results illustrate that the proposed approach correlates reasonably with human judgment and provide an unbiased estimate for the response quality, we believe that there is still room for improvement. For instance, measuring the engagingness of the conversation would be helpful in improving evaluating different dialogue strategies.