PoMo: Generating Entity-Specific Post-Modifiers in Context

We introduce entity post-modifier generation as an instance of a collaborative writing task. Given a sentence about a target entity, the task is to automatically generate a post-modifier phrase that provides contextually relevant information about the entity. For example, for the sentence, "Barack Obama, _______, supported the #MeToo movement.", the phrase "a father of two girls" is a contextually relevant post-modifier. To this end, we build PoMo, a post-modifier dataset created automatically from news articles reflecting a journalistic need for incorporating entity information that is relevant to a particular news event. PoMo consists of more than 231K sentences with post-modifiers and associated facts extracted from Wikidata for around 57K unique entities. We use crowdsourcing to show that modeling contextual relevance is necessary for accurate post-modifier generation. We adapt a number of existing generation approaches as baselines for this dataset. Our results show there is large room for improvement in terms of both identifying relevant facts to include (knowing which claims are relevant gives a >20 BLEU score), and generating appropriate post-modifier text for the context (providing relevant claims is not sufficient for accurate generation). We conduct an error analysis that suggests promising directions for future research.



page 4


The ApposCorpus: A new multilingual, multi-domain dataset for factual appositive generation

News articles, image captions, product reviews and many other texts ment...

Show and Write: Entity-aware News Generation with Image Information

Automatically writing long articles is a complex and challenging languag...

Presenting a New Dataset for the Timeline Generation Problem

The timeline generation task summarises an entity's biography by selecti...

Impact of Target Word and Context on End-to-End Metonymy Detection

Metonymy is a figure of speech in which an entity is referred to by anot...

Automatic Detection of Entity-Manipulated Text using Factual Knowledge

In this work, we focus on the problem of distinguishing a human written ...

Injecting Entity Types into Entity-Guided Text Generation

Recent successes in deep generative modeling have led to significant adv...

Controlled Neural Sentence-Level Reframing of News Articles

Framing a news article means to portray the reported event from a specif...
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

Figure 1: Post-Modifier Generation Task

The goal of machine-in-the-loop writing systems is to assist human writers by directly augmenting their text. Examples include systems that refine human text for grammar Rao and Tetreault (2018), collaborate on story plot generation systems Clark et al. (2018); Yu and Riedl (2012), or modify the content for style Hu et al. (2017); Shen et al. (2017); Yang et al. (2018). In this paper, we introduce post-modifier generation as an instance of such an assistive writing task in the news domain. Journalists use post-modifiers to introduce background information about entities discussed in news articles. To write these post-modifiers journalists often need to look up relevant facts about entities. A post-modifier generation system can be seen as a collaborative assistant that automatically finds relevant facts and inserts a small text fragment that augments the text produced by the human writer.

Post-modifier generation is a contextual

data-to-text generation problem, where the data is the set of known facts about the target entity, and the text to be generated is a post-modifier that is relevant to the rest of the information conveyed in the text. Figure 

1 shows an example. Given a sentence about the anti-war resistance work of Noam Chomsky, the target entity, and a set of known facts about him, the task is to generate a post-modifier that introduces Chomsky as a professor and mentions his background as an anti-war activist. An effective post-modifier generation system must: (i) select suitable facts about the entity given the text, and (ii) produce text that covers these facts in a way that fits in with the rest of the text.

We introduce PoMo, an automatically generated dataset for developing post-modifier generation systems.111https://stonybrooknlp.github.io/PoMo/ PoMo is a collection of sentences that contain entity post-modifiers, along with a collection of facts about the entities obtained from Wikidata Vrandečić and Krötzsch (2014). We use a small number of dependency patterns to automatically identify and extract post-modifiers of entities in sentences. We then link the extracted entities with the entries in Wikidata. The resulting dataset has 231,057 instances covering 57,966 unique entities. Our analysis show that the post-modifiers often combine multiple facts and are specific to the sentential context.

We conduct two sets of experiments that highlight the challenges in post-modifier generation. (i) Claim Selection: Given an input sentence, the first step in generating a post-modifier is to figure out which facts to use. We formulate this as a distantly-supervised ranking problem, where we train neural models that learn to identify relevant claims for a given sentence. These claim ranking models perform well when predicting the relevance of coarse-grained facts (e.g. occupation), but fare poorly when predicting finer-grained facts (e.g. place of birth). (ii) Generation: We adapt recent sequence-to-sequence generation models for this task. Results show that generation remains a challenge. Even though our automatic claim ranking does not improve generation, further experiments with oracle selected claims demonstrate that when relevant claims are known, the models can generate post-modifiers which humans deem comparable in quality to ones written by professional journalists.

In summary, the main contributions of this work are: 1) a data-to-text problem that introduces new challenges, 2) an automated dataset creation pipeline and a large resulting dataset, 3) a crowdsourcing study that verifies the contextual relevance of post-modifiers, and 4) a characterization of the difficulty of the task via performance analysis of numerous baselines.

2 PoMo: Task and Dataset

Post-modifier generation can be formulated as a data-to-text generation problem. The input is text mentioning a target entity and a set of known facts about the entity. The output is a phrase that: (i) fits as a post-modifier of the target entity mentioned in the input text, and (ii) conveys a subset of facts relevant to the context of the input text.

Figure 1 shows an example for the target entity Noam Chomsky. The input includes a sentence mentioning Chomsky’s work on mobilizing anti-war groups along with its surrounding context, and a listing of all facts about Chomsky that are available in Wikidata. Given these inputs, the task is to output a post-modifier phrase that conveys facts about Chomsky that fit within the sentence. In this example the post-modifier conveys both general background information about Chomsky (his occupation), and specific information relevant to the context of the sentence (being an anti-war activist).

This task can be seen as an instance of collaborative writing, where the journalist writes text about specific news events involving entities, and the generation system assists the journalist by inserting new text that augments the story. Given a large collection of news articles, we can automatically create training data for such systems by removing the pieces of text that we want the assistant to generate. This requires reliable ways to identify text to remove and sources of information that can be used to generate the text. Here we describe a pipeline for generating such a dataset for our task.

2.1 Dataset

We construct the PoMo dataset using three different news corpora: NYTimes  Sandhaus (2008), CNN and DailyMail (Hermann et al., 2015). We use Wikidata to collect facts about entities.222Wikidata dump from https://www.wikidata.org/wiki/Wikidata:Database_download (Dump date: 2018/06/25)

Train 6,557 11,323 202,735 220,615
Valid 162 267 4,771 5,200
Test 181 288 4,773 5,242
Total 6,900 11,878 212,279 231,057
Table 1: Dataset distribution by sources.

2.1.1 Post-Modifier and Entity Identification

We use Stanford CoreNLP Manning et al. (2014) to parse each sentence in the news articles and to identify named entities. We extract post-modifiers by finding noun phrases that share an appos relation333An appositional modifier of an NP is another NP immediately to the right that defines or modifies the NP. with any recognized named entity in the sentence. In this work, we only consider post-modifiers for people. In the future, we plan to expand PoMo to include more post-modifiers for other targets, such as organizations. We extract only one such pair from a given sentence to reduce the possible noise in the extraction process.

In our running example from Figure 1, Noam Chomsky is recognized as a person entity. The word “professor” is an appositive dependency of the word “Chomsky” and therefore, we extract the NP “the Massachusetts Institute of Technology professor and antiwar activist” which includes the word “professor” as a post-modifier for the target entity Noam Chomsky.

2.1.2 Entity Claim Matching

Wikidata provides information about entities in the form of key-value pairs that are called claims. To collect the facts about a target entity, we need to link the target to a specific entity in Wikidata. We first search through Wikidata labels and aliases to find candidates with the same name as the target. We sort the candidates based on the number of claims that have a significant word overlap with the extracted post-modifier. We link the entity to the highest ranked candidate whose claims cover at least 30% of the non stop words in the post-modifier. If such a candidate is found we record the claims that overlap with the post-modifier. If no such candidate is found then we discard the entity.

We evaluate this simple heuristic by comparing the results to using an off-the-shelf entity linking system AIDA-light 

Nguyen et al. (2014) and show the results in Table 2. We find that AIDA-light agrees with our entity linking in 91.2% of the cases. AIDA-light is able to link 94.3% of the entities we found from NYTimes, but for CNN and DailyMail, it links only 87.0% and 86.34% of the entities, respectively. This decrease is likely due to the fact that AIDA-light was last updated in 2014 while the CNN/DailyMail datasets contain articles collected until the end of April 2015. On the other hand, NYTimes articles range from 1987 to 2007. Our heuristic seems to be reasonably reliable as it does not depend on anything else but the data sources: news articles and Wikidata.

AIDA Succ. Agreement
Overall 93.66 91.22
Train 93.65 91.16
Valid 94.06 91.13
Test 93.80 93.65
CNN 87.03 90.34
DM 86.34 85.66
NYT 94.29 91.53
Table 2: Percent agreement with AIDA-light’s named entity disambiguation results.

2.2 Analysis

Table 1 shows the distribution of the data sources over train, validation, and test sets. All splits maintain the relative distributions of the data sources to prevent stylistic mismatches from influencing generation. We also ensure that there is no entity overlap among the splits. Within the NYTimes data, we verify that the distribution over years between 1987 and 2007 is also similar over the sets.

Distribution of Post-Modifiers and Entities

Figure 1(a) shows the distribution of post-modifier lengths in terms of token counts. Most post-modifiers are three to eight words long, and about are even longer. Figure 1(b)

shows an estimate of the number of relevant facts covered by the post-modifiers; this estimate uses the number of claims that overlap with the post-modifier via heuristic matching. More than half of the post-modifiers convey two or more facts. About

convey five or more facts. These results suggest that generating post-modifiers requires composing together multiple relevant facts.

(a) Histogram of the token counts of the post-modifiers. Majority of the post-modifiers (171K instances, 74.14%) have 3 to 8 tokens. Average is 5.8 tokens.
(b) Number of relevant facts per instance in the dataset. More than a half of the post-modifiers are related to two or more facts.
(c) Histogram of the scores for post-modifiers, averaged over three annotations. The distribution of ratings for true and other post-modifiers.
Figure 2: PoMo Post-Modifier Statistics

Table 3 lists the most frequent types of facts used in the post-modifiers in our dataset. Most relate to generic biographical information such as the entity’s occupation, organizations they belong to, place of birth, etc. Here again we see a range of types of information being conveyed which is likely to present a challenge for generation systems.

The dataset also covers a wide variety of entity types. We cluster the target entities by their occupation listed in Wikidata. We also use WordNet Miller (1995) to traverse the hypernyms of the words to find frequent ones. Then, we manually select the top ten occupation types. Any entity that does not belong to the top ten is assigned to a single other group. The resulting distribution is shown in Table 4.

Fact Type Count
position held 151,959
occupation 82,781
educated at 53,067
member of political party 42,416
member of sports team 41,602
employer 36,412
award received 31,618
position played on team / speciality 23,987
country of citizenship 17,444
nominated for 15,139
place of birth 9,185
participant of 8,520
member of 7,565
languages spoken, written or signed 4,827
place of death 4,071
Table 3: Top 15 frequent fact types based on heuristic fact coverage identification.
Quality of Post-Modifiers

We conduct a crowdsourcing study to understand how often the post-modifiers are specific to the particular context. For each (entity, context, post-modifier) triple in the validation set, we create multiple alternative post-modifiers by randomly choosing up to ten other post-modifiers that are found in some other sentences for the same entity. Crowd workers rate the quality of these post-modifiers. Figure 3 shows a screenshot of a task given to crowd workers. If the true post-modifier, the one that is actually used in the context, is rated the highest compared to the rest, then we assume the post-modifier is indeed specific to the context. On the other hand, if the crowd workers rate multiple other post-modifiers as good fits for the context, then the true post-modifier is not context specific. Figure 1(c) shows the distribution of ratings for true and other post-modifiers. The true post-modifiers tend to be rated very good or good more often than the other post-modifiers. This suggests that in many cases post-modifiers are specific to the context and cannot be simply replaced by other post-modifiers.

Occupation Count Percentage
athlete 13,560 23.39%
writer 9,177 15.83%
politician 8,518 14.69%
entertainer 6,488 11.19%
other 5,870 10.13%
scientist 4,487 7.74%
artist 4,175 7.20%
official 2,098 3.62%
lawyer 1,132 1.95%
educator 961 1.66%
capitalist 789 1.36%
scholar 711 1.23%
Table 4: Distribution of the inferred occupations of the target entities. Entities clustered by their occupation.
Figure 3: Screenshot of the crowdsourcing task. We asked crowd to rate the quality of post-modifiers.

3 Relevant Claim Selection

One of the key challenges of generating post-modifiers is to identify the claims about an entity that are relevant to the given context. In this section, we explore methods for solving this task.

3.1 Methods

We consider three different models: a most-common claim baseline and two neural baselines.

Most-Common Claim

This model employs a simple frequency heuristic: rank claims by the frequency of their types in the training post-modifiers (e.g. as in the order given in Table 3) and deem the top claims in this ranking as relevant.

Neural Baselines

We use two neural baselines with the following architecture. Word embeddings are used to represent words in the context (e.g. current and previous sentence) and claims. The sequences of embeddings are then fed through 2-layer LSTM’s Hochreiter and Schmidhuber (1997)

to obtain separate representations of the context and claims. These representations are subsequently concatenated together and fed through a fully-connected layer with sigmoid activation, producing a scalar value for each claim representing the probability that it is relevant. We use this model in two ways: as a classifier, and as a ranking model. When used as a classifier, any claim whose score exceeds a threshold

is predicted to be relevant. When used as a ranking model, the top highest-scoring claims are predicted to be relevant.

3.2 Experiments

We train our baselines on the PoMo dataset, using the claims detected during dataset collection as a (distant) source of supervision. Precision, recall, and

score are used to evaluate model performance. Model hyperparameters are chosen using (coarse) grid search to maximize

score on the validation set. The neural baselines use a vocabulary size of 50,000, 100-dimensional word embeddings, and 256 hidden units in the LSTM layers. Dropout Srivastava et al. (2014) is applied between the LSTM layers with a keep probability. The neural classifier uses threshold . We find the optimal value of is for the most-common claims model and for the neural ranker.

Quantitative results are provided in Table 5. Both neural baselines perform considerably better than the most-common claims model. This indicates that the provided contexts and claim values contain useful information for claim selection that goes beyond the information captured by global statistics of the dataset alone. We additionally observe that the ranking-based approach outperforms the classification-based approach in terms of both precision and score, while having only slightly worse recall.

Prec. Recall
Most-Common Claim (=4) 39.9 51.6 45.0
Neural Classifier (=0.37) 52.0 63.8 57.4
Neural Ranker (=2) 66.5 62.7 64.5
Table 5: Baseline model performance on the claim selection task.

To better understand the cases where the neural models fail and succeed, we examine the distribution of scores over the top 15 fact types (see Table 6). Interestingly, when ranked by score we observe that fact types fall naturally into topically related groups:

  1. [nosep]

  2. position / occupation-related facts: position played, position held, occupation

  3. membership-related facts: member of political party, member of, member of sports team

  4. achievement-related facts: award received, nominated for

  5. location-related facts: country of citizenship, place of death, place of birth

With the exception of employer, the overarching trend is that the model identifies the relevance of coarse-grained claims better than fine-grained claims (e.g occupations, political parties, and sports positions are much more likely to be shared between entities than birth and death places). This suggests that developing better methods for determining the relevance of fine-grained claims is a promising avenue for future research on this task.

Fact Type
employer 76.95
position played on team / speciality 76.65
position held 63.10
occupation 50.02
member of political party 48.71
member of 45.60
member of sports team 38.53
award received 37.53
nominated for 30.87
educated at 29.56
participant of 29.04
country of citizenship 16.28
place of death 14.72
place of birth 6.80
languages spoken, written or signed 0.00
Table 6: score of neural ranker () on top 15 fact types.

4 Post-Modifier Generation

We move our focus to the main task of post-modifier generation.

4.1 Methods

At its core, post-modifier generation involves producing a variable-length sequence output conditioned on two variable-length inputs: the words in the current and previous sentence (e.g. the context), and the collection of claims about the entity. Accordingly, the sequence-to-sequence (seq2seq) framework Sutskever et al. (2014) is a natural fit for the task — we use it as the foundation for all of our baseline models. Since research has shown that attention Bahdanau et al. (2015) and copy mechanisms Gu et al. (2016) consistently improve seq2seq model performance, we use these in our baselines as well.

One choice that must be made when using this framework is how to combine the different inputs. The default approach we use is to concatenate the claim and context into a linear sequence of tokens during preprocessing (shown in Figure  3(a)

). We also experiment with encoding the claims and each of the context sentences separately, then concatenating their vector representations before decoding. We refer to this as the

tri-encoder approach (shown in Figure 3(b)).

As discussed earlier, selecting relevant claims is crucial to generating good post-modifiers. One way to incorporate claim selection is to use our baseline models from Section 3 to cut out irrelevant claims from the input before feeding them to the encoder (e.g. performing hard claim selection). This pipelined approach is not differentiable, and can suffer from cascading errors. An alternative way is to use the model’s attention mechanism as a form of soft claim selection that attends only to the relevant claims. The drawback of this approach is that it does not make use of the available claim annotations, which are an important source of supervision.

Building on these observations, we propose an end-to-end claim selection

model which incorporates an additional term to the loss function that encourages the claim-level attention probabilities to be higher for the identified relevant claims as shown in Figure 


. The process for computing this loss term works as follows. We begin by summing together attention scores for tokens within claims to obtain a claim-level score. These scores are then fed through a sigmoid activation function to obtain a soft claim selection probability. For each claim, we measure the binary cross entropy between the predicted selection probability and a binary variable indicating whether or not the claim was identified as relevant. The final loss term is the average of these binary cross entropies. Note that we do not use a copy mechanism in this model to avoid double-counting (since relevant claims were identified using word overlap).

(a) Basic sequence-to-sequence model
(b) Tri-encoder model
(c) End-to-end claim selection model
Figure 4: PoMo Models for post-modifier generation. Grey boxes at the bottom represent individual encoder/decoder modules. (a) For baseline BiLSTM and transformer models all inputs are concatenated into one sequence. (b) The tri-encoder model has a separate encoder and attention for each type of input. The outputs of attention layers are concatenated together before generation. (c) The end-to-end claim selection model attends to only the claim embeddings and uses an auxiliary loss term to encourage high attention scores for relevant claims.

4.2 Experiments

We experiment with two types of encoder/decoder modules: bidirectional LSTMs, and transformers Vaswani et al. (2017). We use a vocabulary of size 50K, truncate the maximum input sequence length to 500, and use a batch size of 32 in all experiments. To help models distinguish between claims and context we demarcate claim fields with special <claim>, <key>, and <value> tokens. We train all the models for 150k steps, and evaluate on the validation dataset every 10k steps. Evaluation is performed using the BLEU Papineni et al. (2002) and METEOR Banerjee and Lavie (2005) translation metrics, and Precision, Recall and score of the predicted bag-of-words (omitting stopwords). The model with the highest score on the validation set is used during test time.

For the bidirectional LSTM, we use 2 hidden layers with 512 hidden units, 500-dimensional word embeddings, and apply dropout between layers with a keep probability of 0.7. Models are trained using stochastic gradient descent with a learning rate of 1.0. For the transformer model, we use 4 attention heads, 4 layers of transformer blocks with 64 hidden units for the encoder and the decoder, a penultimate hidden layer with 256 units, and 64-dimensional word embeddings. Transformer models are trained using Adam 

Kingma and Ba (2015) with an initial learning rate of 2.0, and a label smoothing Szegedy et al. (2016) factor of 0.1 when calculating loss.

We perform a variety of experiments, the results of which are displayed in Table 7. In this table, Transformer and BiLSTM refer to models trained using the default approach to combining context and claims, while Tri-encoder refers to a BiLSTM model trained using the approach described in 4.1 (we do not train a transformer version since its performance is lackluster). Here are detailed descriptions of the experiments performed in each section:

  • All Claims: Results for vanilla seq2seq models.

  • Oracle: Hard claim selection is performed using the oracle relevant claims.

  • Neural Ranker (): Hard claim selection is performed using the top-10 claims returned by the neural ranker baseline.

  • End-to-End Claim Selection: Results for the end-to-end claim selection model.

In order to understand the relative contribution of the different inputs, we also include results for the BiLSTM model trained using either only the claims, or only the context sentences. In Figure 5 and 6, we show the performances by post-modifier and sentence lengths to examine the impact of the such variables.

Discussion of Quantitative Results

Our results contain a few key findings. The first is that knowing the relevant claims is critical to obtaining state-of-the-art performance; even knowing only oracle claims is sufficient to perform better than all of the other baselines, although there is a still a large improvement when context is additionally provided. However, model-based approaches for claim selection do not seem to help: hard claim selection using the neural ranker performs just as well as the vanilla models, and our proposed approach for end-to-end claim selection has a negative impact. This motivates the need for more effective methods of claim selection. The decreasing performances of the BiLSTM seq2seq models by the increasing target post-modifier and sentence lengths show the difficulty of generating long texts and handling long input data. Finally, we observe that the transformer-based seq2seq models are not particularly well-suited to this task. In all cases their performance is inferior to the BiLSTM-based approaches. Large-scale, pre-trained transformer-based language models, such as GPT-2 Radford et al. (2019) and BERT Devlin et al. (2018), might be an interesting addition to the baselines, by framing the task as filling in the blanks for post-modifiers. However, when restricted to approaches that only use our dataset for training, we expect those based on language models to struggle due to the separation of entities among train, validation, and test.

Prec. Rec. BLEU MET.

All Claims
Transformer 41.9 22.2 29.0 7.0 12.1
Tri-Encoder 53.9 32.4 40.5 17.0 17.6
BiLSTM 51.1 34.7 41.4 19.4 18.8

Transformer 69.4 38.6 49.6 15.7 20.0
Tri-Encoder 68.8 47.3 56.1 24.0 24.5
BiLSTM 66.4 48.8 56.2 25.1 25.3

Neural Ranker ()
Transformer 41.5 22.4 29.1 6.9 12.1
Tri-Encoder 53.5 34.1 41.6 17.6 18.3
BiLSTM 49.0 34.2 40.3 18.5 18.5

End-to-End Claim Selection
BiLSTM 47.5 27.9 35.2 13.7 15.3

Context Only
BiLSTM 13.3 8.5 10.3 3.4 6.2

Claims Only
BiLSTM 47.3 28.5 35.6 13.5 15.0

Oracle Claims Only
BiLSTM 63.8 44.7 52.5 21.3 22.7

Table 7: Post modifier generation model performances with seq2seq models. Precision, recall and scores are computed ignoring stopwords.
Input Sky News reported Thursday night that Kenneth Clarke ,           , had not yet decided whether to support Mr. Howard ’s candidacy , raising the possibility the party could face a divisive battle for leadership .
Claims + (position held: Chancellor of the Exchequer)
+ (position held: Secretary of State for the Home Department)
Target a former chancellor of the exchequer
All Claims the Home Secretary
Oracle the Chancellor of the Exchequer
Input “ A lot of people think it ’s something we just started , but we actually opened the season with our first drive using it against Indianapolis , ” said Howard Ballard ,           .
Claims + (member of sports team: Buffalo Bills)
+ (position played on team / speciality: offensive tackle)
+ (mass: 325 pound)
+ (height: 78 inch)
Target Buffalo ’s robust , 6-foot-6-inch , 325-pound right tackle
All Claims & Oracle the Bills ’ offensive tackle
Table 8: Challenging PoMo instances. Two examples along with outputs of the best All Claims and Oracle models are displayed. Claims deemed relevant during dataset curation are prefaced with a +. In the first example, knowing the relevant claims helps the Oracle model produce an output that closely matches the Target, however lack of temporal information causes the model to miss the word former

. In the second example, the All Claims and Oracle models produce the same post-modifier. Although it is similar to the Target in meaning, it receives a low score using our evaluation metrics. Futhermore, our data curation method fails to identify relevant claims.

Qualitative Analysis

A cursory examination of model predictions (see Table 8 for examples) provides insight into why post-modifier generation is a challenging task. One issue that consistently appears is temporal inconsistency between the target and generated post-modifiers. That is, the model may make an error since it is unaware of the time period that the article is written in (and also may not be aware of the periods of time for which a claim are true). For example, in the first instance in Table 8 the Oracle model predicts an almost correct post-modifier but misses the fact that Kenneth Clarke is a former Chancellor of the Exchequer. Another apparent issue is that models tend to generate shorter post-modifiers than humans. As is indicated in Figure 1(a) the post-modifiers in the dataset on average contain 5.8 tokens, whereas generated post-modifiers have only 3.8. Lastly, we observe that our quantitative evaluation metrics can be too strict. Take for example the second instance in Table 8. Here the content of the target and generated post-modifiers is almost exactly the same, however our metrics would give very low scores due to low overlap.

Human Evaluation

We additionally evaluate the generated post-modifiers by performing a human evaluation using Amazon Mechanical Turk. We randomly select 500 instances from test set and show crowdworkers the sentence context, along with the true post-modifier and a generated one. For each instance, workers are asked to select the better phrase, or indicate that the two phrases are of equal quality. For the Oracle BiLSTM model, the true post-modifiers are preferred 46% of the time, while generated post-modifiers are preferred 43.2% of the time. For the Neural Ranker () BiLSTM model, true post-modifiers are favored much more (57.60%) than the generated ones (20%). Consistent with our quantitative results, we see that claim selection is a crucial factor in this task. We also observe a few trends in the results. People tend to prefer generated post-modifiers over the ones written by professional journalists when they are shorter and to use more general terms without elaborating too much about the entity. In contrast, longer and more detailed human written post-modifiers are preferred when they are especially relevant to the rest of the sentence.

(a) scores
(b) BLEU scores
Figure 5: Performances by target post-modifier lengths of BiLSTM model. Post-modifiers with 20 or more tokens are put into one group, .
(a) scores
(b) BLEU scores
Figure 6: Performances by input sentence lengths of BiLSTM model. Sentences with 20 or more tokens are put into one group, .

5 Related Work

There is a large body of previous work on claim selection Kukich (1983); Duboue and McKeown (2003); Reiter and Dale (1997); Tanaka-Ishii et al. (1998); Barzilay and Lapata (2005) and language generation from structured data Reiter et al. (2005); Goldberg et al. (1994). Initially, hand-crafted grammars were employed for language generation, which later evolved to statistical machine translation style models Wong and Mooney (2007) or PCFG based models (Belz, 2008). More recently, the focus has shifted to learning both fact selection and language generation jointly (Liang et al., 2009; Angeli et al., 2010; Kim and Mooney, 2010; Lu and Ng, 2011; Konstas and Lapata, 2013).

Modern approaches employ neural networks to solve this problem end-to-end.

Mei et al. (2016) utilize an encoder-decoder framework to map weather conditions to a weather forecast. Ahn et al. (2016) and Yang et al. (2017) introduce a new class of language models which are capable of entity co-reference and copying facts from an external knowledge base. Building upon these models, Wiseman et al. (2017) introduce an auxiliary reconstruction loss which use the hidden states of the decoder to recover the facts used to generate the text. Liu et al. (2018)

introduce a hierarchical attention model for fact selection, with the higher level focusing on which records in the table to select and the lower level focusing on which cells in a particular row to pay attention to.

In order to train complex neural models, the quest for larger datasets has become paramount. Lebret et al. (2016) introduce the WikiBio dataset containing Wikipedia articles of famous people and the corresponding infobox tables. One drawback of this dataset is that it is easily solved using template-based models. To address this issue, Wiseman et al. (2017) introduce the ROTOWire dataset, which contains summaries of basketball games that are very long and syntactically diverse. A comprehensive list of datasets is provided in Appendix B.

6 Conclusions and Future Work

Inspired by recent work on collaborative writing and data-to-text generation, we introduce post-modifier generation, a task that bridges the gap between these two fields. The task is to generate a factual description of an entity which fits within the context of a human written sentence. In order to promote research on this task we present PoMo

, a large dataset of automatically extracted post-modifiers from news articles, aligned to the Wikidata knowledge graph. We study the performance of numerous strong baseline models on this dataset, with a particular focus on the specific sub-task of claim selection. Our results demonstrate that when relevant claims are known, sequence-to-sequence models are capable of generating post-modifiers which humans deem comparable in quality to ones written by professional journalists. However, according to both quantitative metrics and human judgment, performance is much lower when models must determine for themselves which claims are relevant. These experiments suggest plausible pathways to achieving human-level performance on this task that are both challenging and interesting problems for future research.


We would like to thank the Toyota Technological Institute at Chicago for hosting the Workshop on Collaborative and Knowledge-Backed Language Generation which initiated the efforts for this project. The authors would also like to thank David Yarowsky, Jason Eisner, Kevin Duh, Kyle Gorman, and Philipp Koehn for feedback on early ideas for post-modifier generation.


  • Ahn et al. (2016) Sungjin Ahn, Heeyoul Choi, Tanel Pärnamaa, and Yoshua Bengio. 2016. A neural knowledge language model. arXiv preprint arXiv:1608.00318.
  • Angeli et al. (2010) Gabor Angeli, Percy Liang, and Dan Klein. 2010. A simple domain-independent probabilistic approach to generation. In

    Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing

    , pages 502–512. Association for Computational Linguistics.
  • Bahdanau et al. (2015) Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.
  • Banerjee and Lavie (2005) Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72.
  • Barzilay and Lapata (2005) Regina Barzilay and Mirella Lapata. 2005. Collective content selection for concept-to-text generation. In Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 331–338. Association for Computational Linguistics.
  • Belz (2008) Anja Belz. 2008. Automatic generation of weather forecast texts using comprehensive probabilistic generation-space models. Natural Language Engineering, 14(4):431–455.
  • Clark et al. (2018) Elizabeth Clark, Anne Spencer Ross, Chenhao Tan, Yangfeng Ji, and Noah A Smith. 2018. Creative writing with a machine in the loop: Case studies on slogans and stories. In 23rd International Conference on Intelligent User Interfaces, pages 329–340. ACM.
  • Devlin et al. (2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
  • Duboue and McKeown (2003) Pablo A Duboue and Kathleen R McKeown. 2003. Statistical acquisition of content selection rules for natural language generation. In Proceedings of the 2003 conference on Empirical methods in natural language processing, pages 121–128. Association for Computational Linguistics.
  • Goldberg et al. (1994) Eli Goldberg, Norbert Driedger, and Richard I Kittredge. 1994. Using natural-language processing to produce weather forecasts. IEEE Expert, 9(2):45–53.
  • Gu et al. (2016) Jiatao Gu, Zhengdong Lu, Hang Li, and Victor OK Li. 2016. Incorporating copying mechanism in sequence-to-sequence learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 1631–1640.
  • Hermann et al. (2015) Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems, pages 1693–1701.
  • Hochreiter and Schmidhuber (1997) Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. In Neural Computation.
  • Hu et al. (2017) Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In

    Proceedings of the 34th International Conference on Machine Learning-Volume 70

    , pages 1587–1596. JMLR. org.
  • Kim and Mooney (2010) Joohyun Kim and Raymond J Mooney. 2010. Generative alignment and semantic parsing for learning from ambiguous supervision. In Proceedings of the 23rd International Conference on Computational Linguistics: Posters, pages 543–551. Association for Computational Linguistics.
  • Kingma and Ba (2015) Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.
  • Konstas and Lapata (2013) Ioannis Konstas and Mirella Lapata. 2013. A global model for concept-to-text generation.

    Journal of Artificial Intelligence Research

    , 48:305–346.
  • Kukich (1983) Karen Kukich. 1983. Design of a knowledge-based report generator. In Proceedings of the 21st annual meeting on Association for Computational Linguistics, pages 145–150. Association for Computational Linguistics.
  • Lebret et al. (2016) Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1203–1213.
  • Liang et al. (2009) Percy Liang, Michael I Jordan, and Dan Klein. 2009. Learning semantic correspondences with less supervision. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 1-Volume 1, pages 91–99. Association for Computational Linguistics.
  • Liu et al. (2018) Tianyu Liu, Kexiang Wang, Lei Sha, Baobao Chang, and Zhifang Sui. 2018. Table-to-text generation by structure-aware seq2seq learning. In Thirty-Second AAAI Conference on Artificial Intelligence.
  • Lu and Ng (2011) Wei Lu and Hwee Tou Ng. 2011. A probabilistic forest-to-string model for language generation from typed lambda calculus expressions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 1611–1622. Association for Computational Linguistics.
  • Manning et al. (2014) Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60.
  • Mei et al. (2016) Hongyuan Mei, Mohit Bansal, and Matthew R. Walter. 2016. What to talk about and how? selective generation using lstms with coarse-to-fine alignment. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 720–730, San Diego, California. Association for Computational Linguistics.
  • Miller (1995) George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39–41.
  • Nguyen et al. (2014) Dat Ba Nguyen, Johannes Hoffart, Martin Theobald, and Gerhard Weikum. 2014. Aida-light: High-throughput named-entity disambiguation. LDOW, 1184.
  • Papineni et al. (2002) Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pages 311–318. Association for Computational Linguistics.
  • Radford et al. (2019) Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.
  • Rao and Tetreault (2018) Sudha Rao and Joel R. Tetreault. 2018. Dear sir or madam, may i introduce the gyafc dataset: Corpus, benchmarks and metrics for formality style transfer. In NAACL-HLT.
  • Reiter and Dale (1997) Ehud Reiter and Robert Dale. 1997. Building applied natural language generation systems. Natural Language Engineering, 3(1):57–87.
  • Reiter et al. (2005) Ehud Reiter, Somayajulu Sripada, Jim Hunter, Jin Yu, and Ian Davy. 2005. Choosing words in computer-generated weather forecasts. Artificial Intelligence, 167(1-2):137–169.
  • Sandhaus (2008) Evan Sandhaus. 2008. The new york times annotated corpus. Linguistic Data Consortium, Philadelphia, 6(12):e26752.
  • Shen et al. (2017) Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. In Advances in Neural Information Processing Systems, pages 6830–6841.
  • Srivastava et al. (2014) Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15:1929–1958.
  • Sutskever et al. (2014) Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 3104–3112. Curran Associates, Inc.
  • Szegedy et al. (2016) Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In

    2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016

    , pages 2818–2826.
  • Tanaka-Ishii et al. (1998) Kumiko Tanaka-Ishii, Kôiti Hasida, and Itsuki Noda. 1998. Reactive content selection in the generation of real-time soccer commentary. In Proceedings of the 17th international conference on Computational linguistics-Volume 2, pages 1282–1288. Association for Computational Linguistics.
  • Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc.
  • Vrandečić and Krötzsch (2014) Denny Vrandečić and Markus Krötzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78–85.
  • Wiseman et al. (2017) Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2253–2263, Copenhagen, Denmark. Association for Computational Linguistics.
  • Wong and Mooney (2007) Yuk Wah Wong and Raymond Mooney. 2007. Generation by inverting a semantic parser that uses statistical machine translation. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 172–179.
  • Yang et al. (2017) Zichao Yang, Phil Blunsom, Chris Dyer, and Wang Ling. 2017. Reference-aware language models. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1850–1859, Copenhagen, Denmark. Association for Computational Linguistics.
  • Yang et al. (2018) Zichao Yang, Zhiting Hu, Chris Dyer, Eric P Xing, and Taylor Berg-Kirkpatrick. 2018. Unsupervised text style transfer using language models as discriminators. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages 7287–7298. Curran Associates, Inc.
  • Yu and Riedl (2012) Hong Yu and Mark O. Riedl. 2012. A sequential recommendation approach for interactive personalized story generation. In AAMAS.

Appendix A Additional Claim Selection Materials

Table 9 lists the evaluation results of most-common claim baseline for , the number of claims to predict, from 1 to 5. We obtain highest F1 with .

Precision Recall F1
1 57.7 31.8 41.0
2 49.6 38.3 43.2
3 43.4 45.2 44.3
4 39.9 51.6 45.0
5 36.0 56.8 44.0
Table 9: Evaluation metrics for most-common claim baseline for different values of .

Neural baseline shows improved performance compared to most-common claim baseline, showing its best performance when .

Precision Recall F1
1 75.2 42.4 54.3
2 66.5 62.7 64.5
3 56.0 69.6 62.1
4 48.7 76.2 59.4
Table 10: Evaluation metrics for neural baseline for different values of .
Dataset Size Domain of structured data to language
WEATHER.GOV 29.5k Weather conditions to forecast report
ALLRECIPES 31k Table of ingredients to recipes
ROBOCUP 1.5k Game statistics to summaries
ROTOWIRE 4.9k Basketball statistics to game summaries
WIKIBIO 728k Infobox to Wikipedia biography articles
SBNations 10.9K Game statistic to fan written summaries
WikiFacts 40k Freebase /film/actor facts to Wiki description of actor
Table 11: A comparative analysis of various datasets.

Appendix B Existing Data-to-Text Datasets

Table 11 provides a comprehensive list of data-to-text datasets. PoMo presents a different set of challenges from these datasets. While the target text is shorter and less diverse, the task adds an additional challenge of figuring out which claims to use, a task which our evaluation shows is quite challenging.