MoverScore: Text Generation Evaluating with Contextualized Embeddings and Earth Mover Distance

09/05/2019 ∙ by Wei Zhao, et al. ∙ EPFL Royal Holloway, University of London University of Central Florida Technische Universität Darmstadt 0

A robust evaluation metric has a profound impact on the development of text generation systems. A desirable metric compares system output against references based on their semantics rather than surface forms. In this paper we investigate strategies to encode system and reference texts to devise a metric that shows a high correlation with human judgment of text quality. We validate our new metric, namely MoverScore, on a number of text generation tasks including summarization, machine translation, image captioning, and data-to-text generation, where the outputs are produced by a variety of neural and non-neural systems. Our findings suggest that metrics combining contextualized representations with a distance measure perform the best. Such metrics also demonstrate strong generalization capability across tasks. For ease-of-use we make our metrics available as web service.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

The choice of evaluation metric has a significant impact on the assessed quality of natural language outputs generated by a system. A desirable metric assigns a single, real-valued score to the system output by comparing it with one or more reference texts for content matching. Many natural language generation (NLG) tasks can benefit from robust and unbiased evaluation, including text-to-text (machine translation and summarization), data-to-text (response generation), and image-to-text (captioningGatt and Krahmer (2018). Without proper evaluation, it can be difficult to judge on system competitiveness, hindering the development of advanced algorithms for text generation.

It is an increasingly pressing priority to develop better evaluation metrics given the recent advances in neural text generation. Neural models provide the flexibility to copy content from source text as well as generating unseen words See et al. (2017). This aspect is hardly covered by existing metrics. With greater flexibility comes increased demand for unbiased evaluation. Diversity-promoting objectives make it possible to generate diverse natural language descriptions Li et al. (2016); Wiseman et al. (2018). But standard evaluation metrics including BLEU Papineni et al. (2002) and ROUGE Lin (2004)

compute the scores based primarily on n-gram co-occurrence statistics, which are originally proposed for diagnostic evaluation of systems but not capable of evaluating text quality 

Reiter (2018), as they are not designed to measure if, and to what extent, the system and reference texts with distinct surface forms have conveyed the same meaning. Recent effort on the applicability of these metrics reveals that while compelling text generation system ascend on standard metrics, the text quality of system output is still hard to be improved Böhm et al. (2019).

Our goal in this paper is to devise an automated evaluation metric assigning a single holistic score to any system-generated text by comparing it against human references for content matching. We posit that it is crucial to provide a holistic measure attaining high correlation with human judgments so that various neural and non-neural text generation systems can be compared directly. Intuitively, the metric assigns a perfect score to the system text if it conveys the same meaning as the reference text. Any deviation from the reference content can then lead to a reduced score, e.g., the system text contains more (or less) content than the reference, or the system produces ill-formed text that fails to deliver the intended meaning.

We investigate the effectiveness of a spectrum of distributional semantic representations to encode system and reference texts, allowing them to be compared for semantic similarity across multiple natural language generation tasks. Our new metric quantifies the semantic distance between system and reference texts by harnessing the power of contextualized representations Peters et al. (2018); Devlin et al. (2018) and a powerful distance metric Kusner et al. (2015) for better content matching. Our contributions can be summarized as follows:

  • [topsep=5pt,itemsep=0pt,leftmargin=*]

  • We formulate the problem of evaluating generation systems as measuring the semantic distance between system and reference texts, assuming powerful continuous representations can encode any type of semantic and syntactic deviations.

  • We investigate the effectiveness of existing contextualized representations and Earth Mover’s Distance Rubner et al. (2000) for comparing system predictions and reference texts, leading to our new automated evaluation metric that achieves high correlation with human judgments of text quality.

  • Our metric outperforms or performs comparably to strong baselines on four text generation tasks including summarization, machine translation, image captioning, and data-to-text generation, suggesting this is a promising direction moving forward.

2 Related Work

It is of fundamental importance to design evaluation metrics that can be applied to natural language generation tasks of similar nature, including summarization, machine translation, data-to-text generation, image captioning, and many others. All these tasks involve generating texts of sentence or paragraph length. The system texts are then compared with one or more reference texts of similar length for semantic matching, whose scores indicate how well the systems perform on each task. In the past decades, however, evaluation of these natural language generation tasks has largely been carried out independently within each area.

Summarization  A dominant metric for summarization evaluation is ROUGE Lin (2004), which measures the degree of lexical overlap between a system summary and a set of reference summaries. Its variants consider overlap of unigrams (-1), bigrams (-2), unigrams and skip bigrams with a maximum gap of 4 words (-SU4), longest common subsequences (-L) and its weighted version (-W-1.2), among others. Metrics such as Pyramid Nenkova and Passonneau (2004) and BE Hovy et al. (2006); Tratz and Hovy (2008) further compute matches of content units, e.g., (head-word, modifier) tuples, that often need to be manually extracted from reference summaries. These metrics achieve good correlations with human judgments in the past. However, they are not general enough to account for the relatedness between abstractive summaries and their references, as a system abstract can convey the same meaning using different surface forms. Furthermore, large-scale summarization datasets such as CNN/Daily Mail Hermann et al. (2015) and Newsroom Grusky et al. (2018) use a single reference summary, making it harder to obtain unbiased results when only lexical overlap is considered during summary evaluation.

Machine Translation A number of metrics are commonly used in MT evaluation. Most of these metrics compare system and reference translations based on surface forms such as word/character n-gram overlaps and edit distance, but not the meanings they convey. BLEU Papineni et al. (2002) is a precision metric measuring how well a system translation overlaps with human reference translations using n-gram co-occurrence statistics. Other metrics include SentBLEU, NIST, chrF, TER, WER, PER, CDER, and METEOR Lavie and Agarwal (2007) that are used and described in the WMT metrics shared task Bojar et al. (2017); Ma et al. (2018). RUSE Shimanaka et al. (2018) is a recent effort to improve MT evaluation by training sentence embeddings on large-scale data obtained in other tasks. Additionally, preprocessing reference texts is crucial in MT evaluation, e.g., normalization, tokenization, compound splitting, etc. If not handled properly, different preprocessing strategies can lead to inconsistent results using word-based metrics Post (2018).

Data-to-text Generation BLEU can be poorly suited to evaluating data-to-text systems such as dialogue response generation and image captioning. These systems are designed to generate texts with lexical and syntactic variation, communicating the same information in many different ways. BLEU and similar metrics tend to reward systems that use the same wording as reference texts, causing repetitive word usage that is deemed undesirable to humans Liu et al. (2016). In a similar vein, evaluating the quality of image captions can be challenging. CIDEr Vedantam et al. (2015)

uses tf-idf weighted n-grams for similarity estimation; and SPICE 

Anderson et al. (2016) incorporates synonym matching over scene graphs. Novikova et al. Novikova et al. (2017) examine a large number of word- and grammar-based metrics and demonstrate that they only weakly reflect human judgments of system outputs generated by data-driven, end-to-end natural language generation systems.

Metrics based on Continuous Representations Moving beyond traditional metrics, we envision a new generation of automated evaluation metrics comparing system and reference texts based on semantics rather than surface forms to achieve better correlation with human judgments. A number of previous studies exploit static word embeddings  Ng and Abrecht (2015); Lo (2017) and trained classifers  Peyrard et al. (2017); Shimanaka et al. (2018) to improve semantic similarity estimation, replacing lexical overlaps.

In contemporaneous work, Zhang et al. Zhang et al. (2019) describe a method comparing system and reference texts for semantic similarity leveraging the BERT representations Devlin et al. (2018), which can be viewed as a special case of our metrics and will be discussed in more depth later. More recently, Clark et al. (2019) present a semantic metric relying on sentence mover’s similarity and the ELMo representations  Peters et al. (2018) and apply them to summarization and essay scoring. Mathur et al. (2019) introduce unsupervised and supervised metrics based on the BERT representations to improve MT evaluation, while Peyrard (2019a) provides a composite score combining redundancy, relevance and informativeness to improve summary evaluation.

In this paper, we seek to accurately measure the (dis)similarity between system and reference texts drawing inspiration from contextualized representations and Word Mover’s Distance (WMD; Kusner et al., 2015). WMD finds the “traveling distance” of moving from the word frequency distribution of the system text to that of the reference, which is essential to capture the (dis)similarity between two texts. Our metrics differ from the contemporaneous work in several facets: (i) we explore the granularity of embeddings, leading to two variants of our metric, word mover and sentence mover; (ii) we investigate the effectiveness of diverse pretrained embeddings and finetuning tasks; (iii) we study the approach to consolidate layer-wise information within contextualized embeddings; (iii) our metrics demonstrate strong generalization capability across four tasks, oftentimes outperforming the supervised ones. We now describe our method in detail.

3 Our MoverScore Meric

We have motivated the need for better metrics capable of evaluating disparate NLG tasks. We now describe our metric, namely MoverScore, built upon a combination of (i) contextualized representations of system and reference texts and (ii) a distance between these representations measuring the semantic distance between system outputs and references. It is particularly important for a metric to not only capture the amount of shared content between two texts, i.e., intersect(A,B), as is the case with many semantic textual similarity measures Peters et al. (2018); Devlin et al. (2018); but also to accurately reflect to what extent the system text has deviated from the reference, i.e., union(A,B) - intersect(A,B), which is the intuition behind using a distance metric.

3.1 Measuring Semantic Distance

Let be a sentence viewed as a sequence of words. We denote by the sequence of -grams of (i.e., is the sequence of words and is the sequence of bigrams). Furthermore, let

be a vector of weights, one weight for each

-gram of . We can assume , making a distribution over -grams. Intuitively, the effect of some -grams like those including function words can be downplayed by giving them lower weights, e.g., using Inverse Document Frequency (IDF).

Word Mover’s Distance (WMD) Kusner et al. (2015), a special case of Earth Mover’s Distance Rubner et al. (2000), measures semantic distance between texts by aligning semantically similar words and finding the amount of flow traveling between these words. It was shown useful for text classification and textual similarity tasks Kusner et al. (2015). Here, we formulate a generalization operating on -grams. Let and be two sentences viewed as sequences of -grams: and . If we have a distance metric between -grams, then we can define the transportation cost matrix such that is the distance between the -th -gram of and the -th -gram of . The WMD between the two sequences of -grams and with associated -gram weights and is then given by:

where is the transportation flow matrix with denoting the amount of flow traveling from the -th -gram in to the -th -gram in . Here, denotes the sum of all matrix entries of the matrix , where denotes element-wise multiplication. Then is the minimal transportation cost between and where -grams are weighted by and .

In practice, we compute the Euclidean distance between the embedding representations of -grams: where is the embedding function which maps an -gram to its vector representation. Usually, static word embeddings like word2vec are used to compute but these cannot capture word order or compositionality. Alternatively, we investigate contextualized embeddings like ELMo and BERT because they encode information about the whole sentence into each word vector.

We compute the -gram embeddings as the weighted sum over its word embeddings. Formally, if is the -th -gram from sentence , its embedding is given by:

(1)

where is the IDF of word computed from all sentences in the corpus and is its word vector. Furthermore, the weight associated to the -gram is given by:

(2)

where is a normalizing constant s.t. ,

In the limiting case where is larger than the sentence’s size, contains only one -gram: the whole sentence. Then reduces to computing the distance between the two sentence embeddings, namely Sentence Mover’s Distance (SMD), denoted as:

where and are the size of sentences.

Hard and Soft Alignments

In contemporaneous work, BERTScore (Zhang et al., 2019) also models the semantic distance between system and reference texts for evaluating text generation systems. As shown in Figure 1, BERTScore (precision/recall) can be intuitively viewed as hard alignments (one-to-one) for words in a sentence pair, where each word in one sequence travels to the most semantically similar word in the other sequence. In contrast, MoverScore goes beyond BERTScore as it relies on soft alignments (many-to-one) and allows to map semantically related words in one sequence to the respective word in the other sequence by solving a constrained optimization problem: finding the minimum effort to transform between two texts.

The formulation of Word Mover’s Distance provides an important possibility to bias the metric towards precision or recall by using an asymmetric transportation cost matrix, which bridges a gap between MoverScore and BERTScore:

Figure 1: An illustration of MoverScore and BERTScore.
Proposition 1

BERTScore (precision/recall) can be represented as a (non-optimized) Mover Distance , where is a transportation cost matrix based on BERT and is a uniform transportation flow matrix.222See the proof in the appendix.

3.2 Contextualized Representations

The task formulation naturally lends itself to deep contextualized representations for inducing word vectors . Despite the recent success of multi-layer attentive neural architectures Devlin et al. (2018); Peters et al. (2018), consolidating layer-wise information remains an open problem as different layers capture information at disparate scales and task-specific layer selection methods may be limited Liu et al. (2018, 2019). Tenney et al. (2019) found that a scalar mix of output layers trained from task-dependent supervisions would be effective in a deep transformer-based model. Instead, we investigate aggregation functions to consolidate layer-wise information, forming stationary representations of words without supervision.

Consider a sentence passed through contextualized encoders such as ELMo and BERT with layers. Each layer of the encoders produces a vector representation for each word in . We denote by the representation given by layer , a -dimensional vector. Overall, receives different vectors . An aggregation maps these vectors to one final vector:

(3)

where is the aggregated representation of the word .

We study two alternatives for : (i) the concatenation of power means Rücklé et al. (2018) as a generalized pooling mechanism, and (ii) a routing mechanism for aggregation Zhao et al. (2018, 2019). We relegate the routing method to appendix, as it does not yield better results than power means.

Power Means

Power means is an effective generalization of pooling techniques for aggregating information. It computes a non-linear average of a set of values with an exponent (Eq. (4)). Following Rücklé et al. (2018), we exploit power means to aggregate vector representations pertaining to the -th word from all layers of a deep neural architecture. Let , the -mean of is:

(4)

where exponentiation is applied elementwise. This generalized form can induce common named means such as arithmetic mean (

) and geometric mean (

). In extreme cases, a power mean reduces to the minimum value of the set when , and the maximum value when . The concatenation of -mean vectors we use in this paper is denoted by:

(5)

where is vector concatenation; are exponent values, and we use with in this work.

3.3 Summary of MoverScore Variations

We investigate our MoverScore along four dimensions: (i) the granularity of embeddings, i.e., the size of for -grams, (ii) the choice of pretrained embedding mechanism, (iii) the fine-tuning task used for BERT333ELMo usually requires heavy layers on the top, which restricts the power of fine-tuning tasks for ELMo. (iv) the aggregation technique (-means or routing) when applicable.

Granularity

We used and as well as full sentences ( size of the sentence).

Embedding Mechanism

We obtained word embeddings from three different methods: static embedding with word2vec as well as contextualized embedding with ELMo and BERT. If , -gram embeddings are calculated by Eq. (1). Note that they represent sentence embeddings when size of the sentence.

Fine-tuning Tasks

Natural Language Inference (NLI) and paraphrasing pose high demands in understanding sentence meaning. This motivated us to fine-tune BERT representations on two NLI datasets, MultiNLI and QANLI, and one Paraphrase dataset, QQP—the largest datasets in GLUE Wang et al. (2018). We fine-tune BERT on each of these, yielding different contextualized embeddings for our general evaluation metrics.

Aggregation

For ELMo, we aggregate word representations given by all three ELMo layers, using -means or routing (see the appendix). Word representations in BERT are aggregated from the last five layers, using -means or routing since the representations in the initial layers are less suited for use in downstream tasks Liu et al. (2019).

4 Empirical Evaluation

In this section, we measure the quality of different metrics on four tasks: machine translation, text summarization, image captioning and dialogue generation. Our major focus is to study the correlation between different metrics and human judgment. We employ two text encoders to embed -grams: , which uses a 12-layer transformer, and , which uses a 3-layer BiLSTM. We use Pearson’s and Spearman’s to measure the correlation. We consider two variants of MoverScore: word mover and sentence mover, described below.

Word Mover We denote our word mover notation containing four ingredients as: WMD-Granularity+Embedding+Finetune+Aggregation. For example, WMD-+BERT+MNLI+PMEANS represents the semantic metric using word mover distance where unigram-based word embeddings fine-tuned on MNLI are aggregated by -means.

Sentence Mover We denote our sentence mover notation with three ingredients as: SMD+Embedding+Finetune+Aggregation. For example, SMD+W2V represents the semantic metric using sentence mover distance where two sentence embeddings are computed as the weighted sum over their word2vec embeddings by Eq. (1).

Direct Assessment
Setting Metrics cs-en de-en fi-en lv-en ru-en tr-en zh-en Average
Baselines METEOR++ 0.552 0.538 0.720 0.563 0.627 0.626 0.646 0.610
RUSE(*) 0.624 0.644 0.750 0.697 0.673 0.716 0.691 0.685
BERTScore-F1 0.670 0.686 0.820 0.710 0.729 0.714 0.704 0.719
Sent-Mover Smd + W2V 0.438 0.505 0.540 0.442 0.514 0.456 0.494 0.484
Smd + ELMO + PMeans 0.569 0.558 0.732 0.525 0.581 0.620 0.584 0.595
Smd + BERT + PMeans 0.607 0.623 0.770 0.639 0.667 0.641 0.619 0.652
Smd + BERT + MNLI + PMeans 0.616 0.643 0.785 0.660 0.664 0.668 0.633 0.667
Word-Mover Wmd-1 + W2V 0.392 0.463 0.558 0.463 0.456 0.485 0.481 0.471
Wmd-1 + ELMO + PMeans 0.579 0.588 0.753 0.559 0.617 0.679 0.645 0.631
Wmd-1 + BERT + PMeans 0.662 0.687 0.823 0.714 0.735 0.734 0.719 0.725
Wmd-1 + BERT + MNLI + PMeans 0.670 0.708 0.835 0.746 0.738 0.762 0.744 0.743
Wmd-2 + BERT + MNLI + PMeans 0.679 0.710 0.832 0.745 0.736 0.763 0.740 0.743
Table 1: Absolute Pearson correlations with segment-level human judgments in 7 language pairs on WMT17 dataset.

Baselines

We select multiple strong baselines for each task for comparison: SentBLEU, METEOR++ Guo et al. (2018), and a supervised metric RUSE for machine translation; ROUGE-1 and ROUGE-2 and a supervised metric Peyrard et al. (2017) for text summarization; BLEU and METEOR for dialogue response generation, CIDEr, SPICE, METEOR and a supervised metric LEIC (Cui et al., 2018) for image captioning. We also report BERTScore Zhang et al. (2019) for all tasks (see §2). Due to the page limit, we only compare with the strongest baselines, the rest can be found in the appendix.

4.1 Machine Translation

Data

We obtain the source language sentences, their system and reference translations from the WMT 2017 news translation shared task Bojar et al. (2017). We consider 7 language pairs: from German (de), Chinese (zh), Czech (cs), Latvian (lv), Finnish (fi), Russian (ru), and Turkish (tr), resp. to English. Each language pair has approximately 3,000 sentences, and each sentence has one reference translation and multiple system translations generated by participating systems. For each system translation, at least 15 human assessments are independently rated for quality.

Results

Table 1: In all language pairs, the best correlation is achieved by our word mover metrics that use a BERT pretrained on MNLI as the embedding generator and PMeans to aggregate the embeddings from different BERT layers, i.e., WMD-1/2+BERT+MNLI+PMeans. Note that our unsupervised word mover metrics even outperforms RUSE, a supervised metric. We also find that our word mover metrics outperforms the sentence mover. We conjecture that important information is lost in such a sentence representation while transforming the whole sequence of word vectors into one sentence embedding by Eq. (1).

TAC-2008 TAC-2009
Responsiveness Pyramid Responsiveness Pyramid
Setting Metrics
Baselines (*) 0.715 0.595 0.754 0.652 0.738 0.595 0.842 0.731
ROUGE-1 0.703 0.578 0.747 0.632 0.704 0.565 0.808 0.692
ROUGE-2 0.695 0.572 0.718 0.635 0.727 0.583 0.803 0.694
BERTScore-F1 0.724 0.594 0.750 0.649 0.739 0.580 0.823 0.703
Sent-Mover SMD + W2V 0.583 0.469 0.603 0.488 0.577 0.465 0.670 0.560
SMD + ELMO + PMeans 0.631 0.472 0.631 0.499 0.663 0.498 0.726 0.568
SMD + BERT + PMeans 0.658 0.530 0.664 0.550 0.670 0.518 0.731 0.580
SMD + BERT + MNLI + PMeans 0.662 0.525 0.666 0.552 0.667 0.506 0.723 0.563
Word-Mover Wmd-1 + W2V 0.669 0.549 0.665 0.588 0.698 0.520 0.740 0.647
Wmd-1 + ELMO + PMeans 0.707 0.554 0.726 0.601 0.736 0.553 0.813 0.672
Wmd-1 + BERT + PMeans 0.729 0.595 0.755 0.660 0.742 0.581 0.825 0.690
Wmd-1 + BERT + MNLI + PMeans 0.736 0.604 0.760 0.672 0.754 0.594 0.831 0.701
Wmd-2 + BERT + MNLI + PMeans 0.734 0.601 0.752 0.663 0.753 0.586 0.825 0.694
Table 2: Pearson and Spearman correlations with summary-level human judgments on TAC 2008 and 2009.

4.2 Text Summarization

We use two summarization datasets from the Text Analysis Conference (TAC)444http://tac.nist.gov: TAC-2008 and TAC-2009, which contain 48 and 44 clusters, respectively. Each cluster includes 10 news articles (on the same topic), four reference summaries, and 57 (in TAC-2008) or 55 (in TAC-2009) system summaries generated by the participating systems. Each summary (either reference or system) has fewer than 100 words, and receives two human judgment scores: the Pyramid score Nenkova and Passonneau (2004) and the Responsiveness score. Pyramid measures how many important semantic content units in the reference summaries are covered by the system summary, while Responsiveness measures how well a summary responds to the overall quality combining both content and linguistic quality.

Results

Tables 2: We observe that lexical metrics like ROUGE correlate above-moderate on TAC 2008 and 2009 datasets. In contrast, these metrics perform poorly on other tasks like Dialogue Generation Novikova et al. (2017) and Image Captioning Anderson et al. (2016). Apparently, strict matches on surface forms seems reasonable for extractive summarization datasets. However, we still see that our word mover metrics, i.e., WMD-1+BERT+MNLI+PMeans, perform better than or come close to even the supervised metric .

BAGEL SFHOTEL
Setting Metrics Inf Nat Qual Inf Nat Qual
Baselines BLEU-1 0.225 0.141 0.113 0.107 0.175 0.069
BLEU-2 0.211 0.152 0.115 0.097 0.174 0.071
METEOR 0.251 0.127 0.116 0.111 0.148 0.082
BERTScore-F1 0.267 0.210 0.178 0.163 0.193 0.118
Sent-Mover SMD + W2V 0.024 0.074 0.078 0.022 0.025 0.011
SMD + ELMO + PMeans 0.251 0.171 0.147 0.130 0.176 0.096
SMD + BERT + PMeans 0.290 0.163 0.121 0.192 0.223 0.134
SMD + BERT + MNLI + PMeans 0.280 0.149 0.120 0.205 0.239 0.147
Word-Mover Wmd-1 + W2V 0.222 0.079 0.123 0.074 0.095 0.021
Wmd-1 + ELMO + PMeans 0.261 0.163 0.148 0.147 0.215 0.136
Wmd-1 + BERT + PMeans 0.298 0.212 0.163 0.203 0.261 0.182
Wmd-1 + BERT + MNLI + PMeans 0.285 0.195 0.158 0.207 0.270 0.183
Wmd-2 + BERT + MNLI + PMeans 0.284 0.194 0.156 0.204 0.270 0.182
Table 3: Spearman correlation with utterance-level human judgments for BAGEL and SFHOTEL datasets.

4.3 Data-to-text Generation

We use two task-oriented dialogue datasets: BAGEL Mairesse et al. (2010) and SFHOTEL Wen et al. (2015), which contains 202 and 398 instances of Meaning Representation (MR). Each MR instance includes multiple references, and roughly two system utterances generated by different neural systems. Each system utterance receives three human judgment scores: informativeness, naturalness and quality score Novikova et al. (2017). Informativeness measures how much information a system utterance provides with respect to an MR. Naturalness measures how likely a system utterance is generated by native speakers. Quality measures how well a system utterance captures fluency and grammar.

Results

Tables 3: Interestingly, no metric produces an even moderate correlation with human judgments, including our own. We speculate that current contextualizers are poor at representing named entities like hotels and place names as well as numbers appearing in system and reference texts. However, best correlation is still achieved by our word mover metrics combining contextualized representations.

4.4 Image Captioning

We use a popular image captioning dataset: MS-COCO Lin et al. (2014), which contains 5,000 images. Each image includes roughly five reference captions, and 12 system captions generated by the participating systems from 2015 COCO Captioning Challenge. For the system-level human correlation, each system receives five human judgment scores: M1, M2, M3, M4, M5 (Anderson et al., 2016). The M1 and M2 scores measure overall quality of the captions while M3, M4 and M5 scores measure correctness, detailedness and saliency of the captions. Following Cui et al. (2018), we compare the Pearson correlation with two system-level scores: M1 and M2, since we focus on studying metrics for the overall quality of the captions, leaving metrics understanding captions in different aspects (correctness, detailedness and saliency) to future work.

Results

Table 4: Word mover metrics outperform all baselines except for the supervised metric LEIC, which uses more information by considering both images and texts.

Setting Metric M1 M2
Baselines LEIC(*) 0.939 0.949
METEOR 0.606 0.594
SPICE 0.759 0.750
BERTScore-Recall 0.809 0.749
Sent-Mover SMD + W2V 0.683 0.668
SMD + ELMO + P 0.709 0.712
SMD + BERT + P 0.723 0.747
SMD + BERT + M + P 0.789 0.784
Word-Mover Wmd-1 + W2V 0.728 0.764
Wmd-1 + ELMO + P 0.753 0.775
Wmd-1 + BERT + P 0.780 0.790
Wmd-1 + BERT + M + P 0.813 0.810
Wmd-2 + BERT + M + P 0.812 0.808
Table 4: Pearson correlation with system-level human judgments on MSCOCO dataset. ’M’ and ’P’ are short names.
Metrics cs-en de-en fi-en lv-en
RUSE 0.624 0.644 0.750 0.697
Hmd-F1 + BERT 0.655 0.681 0.821 0.712
Hmd-Recall + BERT 0.651 0.658 0.788 0.681
Hmd-Prec + BERT 0.624 0.669 0.817 0.707
Wmd-unigram + BERT 0.651 0.686 0.823 0.710
Wmd-bigram + BERT 0.665 0.688 0.821 0.712
Table 5: Comparison on hard and soft alignments.

4.5 Further Analysis

Hard and Soft Alignments

 BERTScore is the harmonic mean of BERTScore-Precision and BERTScore-Recall, where both two can be decomposed as a combination of “Hard Mover Distance” (HMD) and BERT (see Prop. 

1).

We use the representations in the 9-th BERT layer for fair comparison of BERTScore and MoverScore and show results on the machine translation task in Table 5. MoverScore outperforms both asymmetric HMD factors, while if they are combined via harmonic mean, BERTScore is on par with MoverScore. We conjecture that BERT softens hard alignments of BERTScore as contextualized embeddings encode information about the whole sentence into each word vector. We also observe that Wmd-bigrams slightly outperforms Wmd-unigrams on 3 out of 4 language pairs.

Distribution of Scores In Figure 2, we take a closer look at sentence-level correlation in MT. Results reveal that the lexical metric SentBLEU can correctly assign lower scores to system translations of low quality, while it struggles in judging system translations of high quality by assigning them lower scores. Our finding agrees with the observations found in Chaganty et al. (2018); Novikova et al. (2017): lexical metrics correlate better with human judgments on texts of low quality than high quality. Peyrard (2019b) further show that lexical metrics cannot be trusted because they strongly disagree on high-scoring system outputs. Importantly, we observe that our word mover metric combining BERT can clearly distinguish texts of two polar qualities.

Figure 2: Score distribution in German-to-English pair.
Figure 3: Correlation in similar language (de-en) and distant language (zh-en) pair, where bordered area shows correlations between human assessment and metrics, the rest shows inter-correlations across metrics and DA is direct assessment rated by language experts.

Correlation Analysis In Figure 3, we observe existing metrics for MT evaluation attaining medium correlations (0.4-0.5) with human judgments but high inter-correlations between themselves. In contrast, our metrics can attain high correlations (0.6-0.7) with human judgments, performing robust across different language pairs. We believe that our improvements come from clearly distinguishing translations that fall on two extremes.

Impact of Fine-tuning Tasks Figure 4 compares Pearson correlations with our word mover metrics combining BERT fine-tuned on three different tasks. We observe that fine-tuning on closely related tasks improves correlations, especially fine-tuning on MNLI leads to an impressive improvement by 1.8 points on average.

Figure 4: Correlation is averaged over 7 language pairs.

4.6 Discussions

We showed that our metric combining contextualized embeddings and Earth Mover’s Distance outperforms strong unsupervised metrics on 3 out of 4 tasks, i.e., METEOR++ on machine translation by 5.7 points, SPICE on image captioning by 3.0 points, and METEOR on dialogue response generation by 2.2 points. The best correlation we achieved is combining contextualized word embeddings and WMD, which even rivals or exceeds SOTA task-dependent supervised metrics across different tasks. Especially in machine translation, our word mover metric pushes correlations in machine translation to 74.3 on average (5.8 points over the SOTA supervised metric and 2.4 points over contemporaneous BERTScore). The major improvements come from contextualized BERT embeddings rather than word2vec and ELMo, and from fine-tuning BERT on large NLI datasets. However, we also observed that soft alignments (MoverScore) marginally outperforms hard alignments (BERTScore). Regarding the effect of -grams in word mover metrics, unigrams slightly outperforms bigrams on average. For the effect of aggregation functions, we suggested effective techniques for layer-wise consolidations, namely -means and routing, both of which are close to the performance of the best layer and on par with each other (see the appendix).

5 Conclusion

We investigated new unsupervised evaluation metrics for text generation systems combining contextualized embeddings with Earth Mover’s Distance. We experimented with two variants of our metric, sentence mover and word mover. The latter has demonstrated strong generalization ability across four text generation tasks, oftentimes even outperforming supervised metrics. Our metric provides a promising direction towards a holistic metric for text generation and a direction towards more ‘human-like’ (Eger et al., 2019) evaluation of text generation systems.

In future work, we plan to avoid the need for costly human references in the evaluation of text generation systems, and instead base evaluation scores on source texts and system predictions only, which would allow for ‘next-level’, unsupervised (in a double sense) and unlimited evaluation (Louis and Nenkova, 2013; Böhm et al., 2019).

Acknowledgments

We thank the anonymous reviewers for their comments, which greatly improved the final version of the paper. This work has been supported by the German Research Foundation as part of the Research Training Group Adaptive Preparation of Information from Heterogeneous Sources (AIPHES) at the Technische Universität Darmstadt under grant No. GRK 1994/1. Fei Liu is supported in part by NSF grant IIS-1909603.

References

  • P. Anderson, B. Fernando, M. Johnson, and S. Gould (2016) SPICE: semantic propositional image caption evaluation. See DBLP:conf/eccv/2016-5, pp. 382–398. External Links: Link, Document Cited by: §2, §4.2, §4.4.
  • F. Böhm, Y. Gao, C. M. Meyer, O. Shapira, I. Dagan, and I. Gurevych (2019) Better rewards yield better summaries: learning to summarise without references. In

    Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing

    ,
    Hong Kong, China. Cited by: §1, §5.
  • O. Bojar, Y. Graham, and A. Kamran (2017) Results of the WMT17 metrics shared task. In Proceedings of the Conference on Machine Translation (WMT), External Links: Link Cited by: §2, §4.1.
  • A. Chaganty, S. Mussmann, and P. Liang (2018) The price of debiasing automatic metrics in natural language evalaution. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 643–653. External Links: Link Cited by: §4.5.
  • E. Clark, A. Celikyilmaz, and N. A. Smith (2019) Sentence mover’s similarity: automatic evaluation for multi-sentence texts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 2748–2760. External Links: Link Cited by: §2.
  • D. Comaniciu and P. Meer (2002) Mean shift: a robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis & Machine Intelligence (5), pp. 603–619. External Links: Link Cited by: §A.2.
  • Y. Cui, G. Yang, A. Veit, X. Huang, and S. Belongie (2018) Learning to evaluate image captioning. In

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    ,
    pp. 5804–5812. External Links: Link Cited by: §4, §4.4.
  • J. Devlin, M. Chang, K. Lee, and K. Toutanova (2018) BERT: pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805. External Links: Link Cited by: §1, §2, §3.2, §3.
  • S. Eger, G. G. Şahin, A. Rücklé, J. Lee, C. Schulz, M. Mesgar, K. Swarnkar, E. Simpson, and I. Gurevych (2019) Text processing like humans do: visually attacking and shielding NLP systems. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Minneapolis, Minnesota, pp. 1634–1647. External Links: Link, Document Cited by: §5.
  • A. Gatt and E. Krahmer (2018) Survey of the state of the art in natural language generation: Core tasks, applications and evaluation.

    Journal of Artificial Intelligence Research (JAIR)

    .
    External Links: Link Cited by: §1.
  • M. Grusky, M. Naaman, and Y. Artzi (2018) NEWSROOM: a dataset of 1.3 million summaries with diverse extractive strategies. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), External Links: Link Cited by: §2.
  • Y. Guo, C. Ruan, and J. Hu (2018) Meteor++: incorporating copy knowledge into machine translation evaluation. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pp. 740–745. External Links: Link Cited by: §4.
  • K. M. Hermann, T. Kocisky, E. Grefenstette, L. Espeholt, W. Kay, M. Suleyman, and P. Blunsom (2015) Teaching machines to read and comprehend. In Proceedings of Neural Information Processing Systems (NIPS), External Links: Link Cited by: §2.
  • E. Hovy, C. Lin, L. Zhou, and J. Fukumoto (2006) Automated summarization evaluation with basic elements. In Proceedings of the Fifth Conference on Language Resources and Evaluation (LREC 2006), pp. 604–611. External Links: Link Cited by: §2.
  • M. J. Kusner, Y. Sun, N. I. Kolkin, and K. Q. Weinberger (2015) From word embeddings to document distances. In

    Proceedings of the International Conference on Machine Learning (ICML)

    ,
    External Links: Link Cited by: §1, §2, §3.1.
  • A. Lavie and A. Agarwal (2007) Meteor: An Automatic Metric for MT Evaluation with High Levels of Correlation with Human Judgments. In Proceedings of the Second Workshop on Statistical Machine Translation, StatMT ’07, Stroudsburg, PA, USA, pp. 228–231. External Links: Link Cited by: §2.
  • J. Li, M. Galley, C. Brockett, J. Gao, and B. Dolan (2016) A diversity-promoting objective function for neural conversation models. In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL), External Links: Link Cited by: §1.
  • C. Lin (2004) ROUGE: A Package for Automatic Evaluation of summaries. In Proceedings of ACL workshop on Text Summarization Branches Out, Barcelona, Spain, pp. 74–81. External Links: Link Cited by: §1, §2.
  • T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick (2014) Microsoft coco: common objects in context. In European conference on computer vision, pp. 740–755. External Links: Link Cited by: §4.4.
  • C. Liu, R. Lowe, I. V. Serban, M. Noseworthy, L. Charlin, and J. Pineau (2016) How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), External Links: Link Cited by: §2.
  • L. Liu, X. Ren, J. Shang, X. Gu, J. Peng, and J. Han (2018) Efficient contextualized representation: Language model pruning for sequence labeling. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), External Links: Link Cited by: §3.2.
  • N. F. Liu, M. Gardner, Y. Belinkov, M. Peters, and N. A. Smith (2019) Linguistic knowledge and transferability of contextual representations. arXiv preprint arXiv:1903.08855. External Links: Link Cited by: §3.2, §3.3.
  • C. Lo (2017) MEANT 2.0: accurate semantic MT evaluation for any output language. See DBLP:conf/wmt/2017, pp. 589–597. External Links: Link Cited by: §2.
  • A. Louis and A. Nenkova (2013) Automatically assessing machine summary content without a gold standard. Computational Linguistics 39 (2), pp. 267–300. External Links: Link, Document Cited by: §5.
  • Q. Ma, O. Bojar, and Y. Graham (2018) Results of the WMT18 metrics shared task. In Proceedings of the Third Conference on Machine Translation (WMT), External Links: Link Cited by: §2.
  • F. Mairesse, M. Gašić, F. Jurčíček, S. Keizer, B. Thomson, K. Yu, and S. Young (2010)

    Phrase-based statistical language generation using graphical models and active learning

    .
    In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pp. 1552–1561. External Links: Link Cited by: §4.3.
  • N. Mathur, T. Baldwin, and T. Cohn (2019) Putting evaluation in context: contextual embeddings improve machine translation evaluation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 2799–2808. External Links: Link Cited by: §2.
  • A. Nenkova and R. J. Passonneau (2004) Evaluating content selection in summarization: the pyramid method. In Proceedings of the 2004 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 145–152. External Links: Link Cited by: §2, §4.2.
  • J. Ng and V. Abrecht (2015) Better summarization evaluation with word embeddings for rouge. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, Lisbon, Portugal, pp. 1925–1930. External Links: Link Cited by: §2.
  • J. Novikova, O. Dušek, A. Cercas Curry, and V. Rieser (2017) Why We Need New Evaluation Metrics for NLG. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, pp. 2241–2252. External Links: Link, Link, Document Cited by: §2, §4.2, §4.3, §4.5.
  • K. Papineni, S. Roukos, T. Ward, and W. Zhu (2002) BLEU: A Method for Automatic Evaluation of Machine Translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, Stroudsburg, PA, USA, pp. 311–318. External Links: Link, Document Cited by: §1, §2.
  • M. E. Peters, M. Neumann, M. Iyyer, M. Gardner, C. Clark, K. Lee, and L. Zettlemoyer (2018) Deep contextualized word representations. In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL), External Links: Link Cited by: §1, §2, §3.2, §3.
  • M. Peyrard, T. Botschen, and I. Gurevych (2017) Learning to score system summaries for better content selection evaluation. In Proceedings of the Workshop on New Frontiers in Summarization, External Links: Link Cited by: §2, §4.
  • M. Peyrard (2019a) A simple theoretical model of importance for summarization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 1059–1073. External Links: Link Cited by: §2.
  • M. Peyrard (2019b) Studying summarization evaluation metrics in the appropriate scoring range. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 5093–5100. External Links: Link Cited by: §4.5.
  • M. Post (2018) A call for clarity in reporting bleu scores. In Proceedings of the Third Conference on Machine Translation (WMT), External Links: Link Cited by: §2.
  • E. Reiter (2018) A structured review of the validity of BLEU. Computational Linguistics 44 (3), pp. 393–401. External Links: Link, Document Cited by: §1.
  • Y. Rubner, C. Tomasi, and L. J. Guibas (2000)

    The earth mover’s distance as a metric for image retrieval

    .
    International Journal of Computer Vision. External Links: Link Cited by: 2nd item, §3.1.
  • A. Rücklé, S. Eger, M. Peyrard, and I. Gurevych (2018) Concatenated power mean word embeddings as universal cross-lingual sentence representations. CoRR abs/1803.01400. External Links: Link Cited by: §3.2, §3.2.
  • A. See, P. J. Liu, and C. D. Manning (2017) Get to the point: Summarization with pointer-generator networks. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), External Links: Link Cited by: §1.
  • H. Shimanaka, T. Kajiwara, and M. Komachi (2018) RUSE: regressor using sentence embeddings for automatic machine translation evaluation. In Proceedings of the Third Conference on Machine Translation (WMT), External Links: Link Cited by: §2, §2.
  • I. Tenney, P. Xia, B. Chen, A. Wang, A. Poliak, R. T. McCoy, N. Kim, B. Van Durme, S. R. Bowman, D. Das, et al. (2019) What do you learn from context? probing for sentence structure in contextualized word representations. arXiv preprint arXiv:1905.06316. External Links: Link Cited by: §3.2.
  • S. Tratz and E. H. Hovy (2008) Summarization Evaluation Using Transformed Basic Elements. In Proceedings of the text analysing conference, (TAC 2008), External Links: Link Cited by: §2.
  • R. Vedantam, C. L. Zitnick, and D. Parikh (2015) CIDEr: Consensus-based Image Description Evaluation. See DBLP:conf/cvpr/2015, pp. 4566–4575. External Links: Link, Document Cited by: §2.
  • M. P. Wand and M. C. Jones (1994) Kernel smoothing. Chapman and Hall/CRC. External Links: Link Cited by: §A.2.
  • A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. R. Bowman (2018) Glue: a multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461. External Links: Link Cited by: §3.3.
  • T. Wen, M. Gasic, N. Mrksic, P. Su, D. Vandyke, and S. Young (2015) Semantically conditioned lstm-based natural language generation for spoken dialogue systems. arXiv preprint arXiv:1508.01745. External Links: Link Cited by: §4.3.
  • S. Wiseman, S. M. Shieber, and A. M. Rush (2018) Learning neural templates for text generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), External Links: Link Cited by: §1.
  • S. Zhang, W. Zhao, X. Wu, and Q. Zhou (2018)

    Fast dynamic routing based on weighted kernel density estimation

    .
    CoRR abs/1805.10807. External Links: Link, 1805.10807 Cited by: §A.2.
  • T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi (2019) BERTScore: evaluating text generation with BERT. CoRR abs/1904.09675. External Links: Link, 1904.09675 Cited by: §2, §3.1, §4.
  • W. Zhao, H. Peng, S. Eger, E. Cambria, and M. Yang (2019) Towards scalable and reliable capsule networks for challenging NLP applications. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 1549–1559. External Links: Link Cited by: §3.2.
  • W. Zhao, J. Ye, M. Yang, Z. Lei, S. Zhang, and Z. Zhao (2018) Investigating capsule networks with dynamic routing for text classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, External Links: Document, Link Cited by: §3.2.

Appendix A Supplemental Material

a.1 Proofs

In this section, we prove the Prop. 1 in the paper about viewing BERTScore (precision/recall) as a (non-optimized) Mover Distance.

As a reminder, WMD formulates as:

where and denote a vector of weights for each -gram of and .

The BERTScore is defined as:

Then, can be formulated in a “quasi” WMD form:

where and is the size of -grams in . Similarly, we can have in a quasi WMD form (ignored). Then, can be formulated as harmonic-mean of two WMD forms of and .

a.2 Routing

In this section, we study the aggregation function with routing scheme, which achieved promising results on NLP tasks. Specifically, we introduce a nonparametric clustering with Kernel Density Estimation (KDE) for routing since KDE bridges a family of kernel functions with underlying empirical distributions, which often leads to computational efficiency Zhang et al. (2018), defined as:

where is a distance function, denotes underlying closeness between the aggregated vector and vector in the -th layer, some instances of Wand and Jones (1994) can be illustrated as:

One typical solution for KDE clustering to minimize is taking Mean Shift Comaniciu and Meer (2002), defined as:

Firstly, can be updated while is fixed:

Intuitively, can be explained as a final aggregated vector from contextualized layers. Then, we adopt SGD to update :

where is the hyper parameter to control step size. The routing process is summarized in Algorithm 1.

1:procedure ROUTING(, )
2:Initialize
3:while true do
4:   foreach representation and in layer and  do    
5:   foreach representation in layer  do
6:          
7:   foreach representation and in layer and  do    
8:   
9:   if  then
10:      break
11:   else
12:          
13:return
Algorithm 1 Aggregation by Routing

Best Layer and Layer-wise Consolidation

Table 6 compares our Word Mover based metric combining BERT representations on different layers with stronger BERT representations consolidated from these layers (-means and routing). We often see that which layer has best performance is task-dependent, and our WMD-based metrics with -means or routing schema come close to the oracle performance obtained from best layers.

Direct Assessment
Metrics cs-en de-en fi-en lv-en ru-en tr-en zh-en
Wmd-1 + BERT + Layer 8 .6361 .6755 .8134 .7033 .7273 .7233 .7175
Wmd-1 + BERT + Layer 9 .6510 .6865 .8240 .7107 .7291 .7357 .7195
Wmd-1 + BERT + Layer 10 .6605 .6948 .8231 .7158 .7363 .7317 .7168
Wmd-1 + BERT + Layer 11 .6695 .6845 .8192 .7048 .7315 .7276 .7058
Wmd-1 + BERT + Layer 12 .6677 .6825 .8194 .7188 .7326 .7291 .7064
Wmd-1 + BERT + Routing .6618 .6897 .8225 .7122 .7334 .7301 .7182
Wmd-1 + BERT + PMeans .6623 .6873 .8234 .7139 .7350 .7339 .7192
Table 6: Absolute Pearson correlations with segment-level human judgments on WMT17 to-English translations.

Experiments

Table  7, 8 and 9 show correlations between metrics (all baseline metrics and Word Mover based metrics) and human judgments on machine translation, text summarization and dialogue response generation tasks.

Direct Assessment
Setting Metrics cs-en de-en fi-en lv-en ru-en tr-en zh-en Average
Baselines BLEND 0.594 0.571 0.733 0.594 0.622 0.671 0.661 0.635
RUSE 0.624 0.644 0.750 0.697 0.673 0.716 0.691 0.685
SentBLEU 0.435 0.432 0.571 0.393 0.484 0.538 0.512 0.481
chrF++ 0.523 0.534 0.678 0.520 0.588 0.614 0.593 0.579
METEOR++ 0.552 0.538 0.720 0.563 0.627 0.626 0.646 0.610
BERTScore-F1 0.670 0.686 0.820 0.710 0.729 0.714 0.704 0.719
Word-Mover Wmd-1 + W2V 0.392 0.463 0.558 0.463 0.456 0.485 0.481 0.471
Wmd-1 + BERT + Routing 0.658 0.689 0.823 0.712 0.733 0.730 0.718 0.723
Wmd-1 + BERT + MNLI + Routing 0.665 0.705 0.834 0.744 0.735 0.752 0.736 0.739
Wmd-2 + BERT + MNLI + Routing 0.676 0.706 0.831 0.743 0.734 0.755 0.732 0.740
Wmd-1 + BERT + PMeans 0.662 0.687 0.823 0.714 0.735 0.734 0.719 0.725
Wmd-1 + BERT + MNLI + PMeans 0.670 0.708 0.835 0.746 0.738 0.762 0.744 0.743
Wmd-2 + BERT + MNLI + PMeans 0.679 0.710 0.832 0.745 0.736 0.763 0.740 0.743
Table 7: Absolute Pearson correlations with segment-level human judgments on WMT17 to-English translations.
TAC-2008 TAC-2009
Responsiveness Pyramid Responsiveness Pyramid
Setting Metrics
Baselines 0.696 0.558 0.753 0.652 0.731 0.552 0.838 0.724
0.715 0.595 0.754 0.652 0.738 0.595 0.842 0.731
TFIDF-1 0.176 0.224 0.183 0.237 0.187 0.222 0.242 0.284
TFIDF-2 0.047 0.154 0.049 0.182 0.047 0.167 0.097 0.233
ROUGE-1 0.703 0.578 0.747 0.632 0.704 0.565 0.808 0.692
ROUGE-2 0.695 0.572 0.718 0.635 0.727 0.583 0.803 0.694
ROUGE-1-WE 0.571 0.450 0.579 0.458 0.586 0.437 0.653 0.516
ROUGE-2-WE 0.566 0.397 0.556 0.388 0.607 0.413 0.671 0.481
ROUGE-L 0.681 0.520 0.702 0.568 0.730 0.563 0.779 0.652
Frame-1 0.658 0.508 0.686 0.529 0.678 0.527 0.762 0.628
Frame-2 0.676 0.519 0.691 0.556 0.715 0.555 0.781 0.648
BERTScore-F1 0.724 0.594 0.750 0.649 0.739 0.580 0.823 0.703
Word-Mover Wmd-1 + W2V 0.669 0.559 0.665 0.611 0.698 0.520 0.740 0.647
Wmd-1 + BERT + Routing 0.729 0.601 0.763 0.675 0.740 0.580 0.831 0.700
Wmd-1 + BERT + MNLI + Routing 0.734 0.609 0.768 0.686 0.747 0.589 0.837 0.711
Wmd-2 + BERT + MNLI + Routing 0.731 0.593 0.755 0.666 0.753 0.583 0.827 0.698
Wmd-1 + BERT + PMeans 0.729 0.595 0.755 0.660 0.742 0.581 0.825 0.690
Wmd-1 + BERT + MNLI + PMeans 0.736 0.604 0.760 0.672 0.754 0.594 0.831 0.701
Wmd-2 + BERT + MNLI + PMeans 0.734 0.601 0.752 0.663 0.753 0.586 0.825 0.694
Table 8: Correlation of automatic metrics with summary-level human judgments for TAC-2008 and TAC-2009.
BAGEL SFHOTEL
Setting Metrics Inf Nat Qual Inf Nat Qual
Baselines BLEU-1 0.225 0.141 0.113 0.107 0.175 0.069
BLEU-2 0.211 0.152 0.115 0.097 0.174 0.071
BLEU-3 0.191 0.150 0.109 0.089 0.161 0.070
BLEU-4 0.175 0.141 0.101 0.084 0.104 0.056
ROUGE-L 0.202 0.134 0.111 0.092 0.147 0.062
NIST 0.207 0.089 0.056 0.072 0.125 0.061
CIDEr 0.205 0.162 0.119 0.095 0.155 0.052
METEOR 0.251 0.127 0.116 0.111 0.148 0.082
BERTScore-F1 0.267 0.210 0.178 0.163 0.193 0.118
Word-Mover Wmd-1 + W2V 0.222 0.079 0.123 0.074 0.095 0.021
Wmd-1 + BERT + Routing 0.294 0.209 0.156 0.208 0.256 0.178
Wmd-1 + BERT + MNLI + Routing 0.278 0.180 0.144 0.211 0.252 0.175
Wmd-2 + BERT + MNLI + Routing 0.279 0.182 0.147 0.204 0.252 0.172
Wmd-1 + BERT + PMeans 0.298 0.212 0.163 0.203 0.261 0.182
Wmd-1 + BERT + MNLI + PMeans 0.285 0.195 0.158 0.207 0.270 0.183
Wmd-2 + BERT + MNLI + PMeans 0.284 0.194 0.156 0.204 0.270 0.182
Table 9: Spearman correlation with utterance-level human judgments for BAGEL and SFHOTEL datasets.