Fine-grained Information Status Classification Using Discourse Context-Aware Self-Attention

TL;DR

This paper introduces a discourse context-aware self-attention neural network utilizing BERT for fine-grained information status classification, achieving state-of-the-art results and improving bridging anaphora recognition without relying on hand-crafted features.

Contribution

It presents a novel neural model that leverages discourse context and BERT representations for improved information status classification and bridging anaphora recognition.

Findings

Achieves 4.1% accuracy improvement over previous methods.

Improves F1 score for bridging anaphora recognition by 3.9%.

Eliminates need for complex hand-crafted semantic features.

Abstract

Previous work on bridging anaphora recognition (Hou et al., 2013a) casts the problem as a subtask of learning fine-grained information status (IS). However, these systems heavily depend on many hand-crafted linguistic features. In this paper, we propose a discourse context-aware self-attention neural network model for fine-grained IS classification. On the ISNotes corpus (Markert et al., 2012), our model with the contextually-encoded word representations (BERT) (Devlin et al., 2018) achieves new state-of-the-art performances on fine-grained IS classification, obtaining a 4.1% absolute overall accuracy improvement compared to Hou et al. (2013a). More importantly, we also show an improvement of 3.9% F1 for bridging anaphora recognition without using any complex hand-crafted semantic features designed for capturing the bridging phenomenon.

Tables

Table 1. Table 1: Information status categories and their main affecting factors. “Local context” means the sentence s 𝑠 s which contains the target mention, “Previous context” indicates all sentences from the discourse which occur before s 𝑠 s .

			Factors affecting IS
	Description	Example	Mention	Local	Previous
			Properties	Context	Context
old	coreferent with an already introduced entity	he, the president	✓	✓	✓
m/worldKnow.	generally known to the hearer	Francis, the pope	✓	✓
m/syntactic	syntactically linked to other old or mediated	their father	✓
	mentions	a war in Africa
m/aggregate	coordinated NPs where at least one element	U.S. and Canada	✓
	is old or mediated	he and his son
m/function	refer to a value of a previously explicitly	(the price went	✓	✓
	mentioned rise/fall function	down) 6 cents
m/comparative	usually contain a premodifier to indicate that	another law	✓	✓	✓
	this entity is compared to another entity	further attacks
m/bridging	associative anaphors which link to previously	the price	✓	✓	✓
	introduced related entities/events	the reason
new	introduced into the discourse for the first time	a reader	✓	✓
	and not known to the hearer before	politics

Table 2. Table 2: IS distribution in ISNotes.

Mentions		10,980
old		3237	29.5%
mediated		3,708	33.8%
	syntactic	1,592	14.5%
	world knowledge	924	8.4%
	bridging	663	6.0%
	comparative	253	2.3%
	aggregate	211	1.9%
	func	65	0.6%
new		4,035	36.7%

Table 3. Table 3: Results of the discourse context-aware self-attention model compared to the baselines. Bolded scores indicate the best performance for each IS class. The improvements of self-attention with context I and self-attention with context II over the baselines are statistically significant at p < 0.01 𝑝 0.01 p<0.01 using randomization test.

	collective			cascade collective			self-attention			self-attention			self-attention
	Hou et al.(2013)			Hou et al.(2013)			wo context			with context I			with context II
	R	P	F	R	P	F	R	P	F	R	P	F	R	P	F
old	84.4	86.0	85.2	82.2	87.2	84.7	79.3	82.4	80.8	80.1	85.8	82.9	86.8	90.9	88.8
m/worldKnow.	67.4	77.3	72.0	67.2	77.2	71.9	57.8	62.2	59.9	60.8	64.6	62.7	74.6	79.6	77.0
m/syntactic	82.2	81.9	82.0	81.6	82.5	82.0	83.7	81.7	82.7	83.2	82.6	82.9	83.6	83.0	83.3
m/aggregate	64.5	79.5	71.2	63.5	77.9	70.0	77.7	75.2	76.5	76.3	79.7	78.0	76.8	80.6	78.6
m/function	67.7	72.1	69.8	67.7	72.1	69.8	42.2	56.3	48.2	54.7	61.4	57.9	50.0	74.4	59.8
m/comparative	81.8	82.1	82.0	86.6	78.2	82.2	89.3	88.6	89.0	90.1	86.4	88.2	91.7	89.2	90.4
m/bridging	19.3	39.0	25.8	44.9	39.8	42.2	37.6	48.6	42.4	43.6	51.6	47.3	42.5	50.4	46.1
new	86.5	76.1	81.0	83.0	78.1	80.5	86.1	80.0	82.9	88.3	81.0	84.5	88.5	82.0	85.1
acc	78.9			78.6			78.1			79.8			83.0

Full text

1 Introduction

Information Structure Halliday (1967); Prince (1981, 1992); Gundel et al. (1993); Lambrecht (1994); Birner and Ward (1998); Kruijff-Korbayová and Steedman (2003) studies structural and semantic properties of a sentence according to its relation to the discourse context. Information structure affects how discourse entities are referred to in a text, which is known as Information Status Halliday (1967); Prince (1981); Nissim et al. (2004). Specifically, information status (IS henceforth) reflects the accessibility of a discourse entity based on the evolving discourse context and the speaker’s assumption about the hearer’s knowledge and beliefs. For instance, according to Markert et al. (2012), old mentions111A mention is a noun phrase which refers to a discourse entity and carries information status. refer to entities that have been referred to previously; mediated mentions have not been mentioned before but are accessible to the hearer by reference to another old mention or to prior world knowledge; and new mentions refer to entities that are introduced to the discourse for the first time and are not known to the hearer before.

In this paper, we follow the IS scheme proposed by Markert et al. (2012) and focus on learning fine-grained IS on written texts. A mention’s semantic and syntactic properties can signal its information status. For instance, indefinite NPs tend to be new and pronouns are likely to be old. Moreover, referential patterns of how a mention is referred to in a sentence also affect this mention’s IS. In Example 1, “Friends” is a bridging anaphor even if we do not know the antecedent (i.e., she); while the information status for “Friends” in Example 1 is mediated/worldKnowledge. Section 3.1 analyzes the characteristics of each IS category and the relations between IS and discourse context.

• She made money, but spent more. Friends pitched in.

Friends are part of the glue that holds life and faith together.

In this work, we propose a discourse context-aware self-attention neural network model for fine-grained IS classification. We find that the sentence containing the target mention as well as the lexical overlap information between the target mention and the preceding mentions are the most important discourse context when assigning IS for a mention. With self-attention, our model can capture important signals within a mention and the interactions between the mention and its context. On the ISNotes corpus Markert et al. (2012), our model with the contextually-encoded word representations (BERT) Devlin et al. (2018) achieves new state-of-the-art performances on fine-grained IS classification, obtaining a 4.1% absolute overall accuracy improvement compared to Hou et al. (2013a). More importantly, we also show an improvement of 3.9% F1 for bridging anaphora recognition without using any sophisticated hand-crafted semantic features.

2 Related Work

IS classification and bridging anaphora recognition.

Bridging resolution Hou et al. (2014, 2018) contains two sub tasks: identifying bridging anaphors Markert et al. (2012); Hou et al. (2013a); Hou (2016) and finding the correct antecedent among candidates Hou et al. (2013b); Hou (2018a, b). Previous work handle bridging anaphora recognition as part of IS classification problem. Markert et al. (2012) applied joint inference for IS classification on the ISNotes corpus but reported very low results on bridging recognition. Built on this work, Hou et al. (2013a) designed many linguistic features to capture bridging and integrated them into a cascading collective classification algorithm. Differently, Hou (2016) used an attention-based LSTM model based on GloVe vectors and a small set of features for IS classification. The author reported similar results as Hou et al. (2013a) regarding the overall IS classificiation accuracy but the result on bridging anaphora recognition is much worse than Hou et al. (2013a).

Rahman and Ng (2012) incorporated carefully designed rules into an SVMmulticlass algorithm for IS classification on the Switchboard dialogue corpus Nissim et al. (2004). Cahill and Riester (2012) trained a CRF model with syntactic and surface features for fine-grained IS classification on the German DIRNDL radio news corpus Riester et al. (2010)222Bridging antecedent information was not annotated in Switchboard and DIRNDL. Also IS annotation in Switchboard includes non-anaphoric cases..

Different from the above mentioned work, we do not use any complicated hand-crafted features and our model improves the previous state-of-the-art results on both overall IS classification accuracy and bridging recognition by a large margin on the ISNotes corpus.

Self-attention.

Recently, multi-head self-attention encoder Ashish et al. (2017) has been shown to perform well in various NLP tasks, including semantic role labelling Strubell et al. (2018), question answering and natural language inference Devlin et al. (2018). In our model, we create a “pseudo sentence” for each mention and apply the transformer encoder for our task. The self-attention mechanism allows our model to attend to both the context and the mention itself for clues which are helpful for predicting the mention’s IS.

Fine-tuning with contextual word embeddings.

Recent work Peters et al. (2018); Howard and Ruder (2018); Devlin et al. (2018) have shown that a range of downstream NLP tasks benefit from fine-tuning task-specific parameters with pre-trained contextual word representations. Our work belongs to this category and we fine-tune our model based on BERTBASE representations Devlin et al. (2018).

3 Approach

3.1 Information Status and Discourse Context

The IS scheme proposed by Markert et al. (2012) adopts three major IS categories (old, new and mediated) from Nissim et al. (2004) and distinguishes six subcategories for mediated. Table 1 lists the definitions for these IS categories and summarizes the main affecting factors for each IS class.

As described in Section 1, a mention’s internal syntactic and semantic properties can signal its IS. For instance, a mention containing a possessive pronoun modifier is likely to be mediated/syntactic (e.g., their father); and a mediated/comparative mention often contains a premodifier indicating that this entity is compared to another preceding entity (e.g., further attacks).

In addition, for some IS classes, the “local context” (the sentence $s$ which contains the target mention) and “previous context” (sentences from the discourse which precede $s$) play an important role when assigning IS to a mention. Example 1 and Example 1 in Section 1 demonstrate the role of the local context for IS. And sometimes we need to look at the previous context when deciding IS for a mention. In Example 3.1, without looking at the previous context, we tend to think the IS for “Poland” in the second sentence is mediated/WorldKnowledge. Here the correct IS for “Poland” is old because it is mentioned before in the previous context.

• [Previous context:] In Poland, only 4% of all investment goes toward making things farmers want; in the West, it is closer to 20%.

[Local context:] A private farmer in Poland is free to buy and sell land.

3.2 IS Classification with Discourse Context-Aware Self-Attention

To account for the different factors described in the previous section when predicting IS for a mention, we create a novel “pseudo sentence” for each mention and apply the multi-head self-attention encoder Ashish et al. (2017) for this sentence.

Figure 1 depicts the high-level structure of our model. The pseudo sentence consists of five parts: previous overlap_info, local context, the delimiter token “[delimiter]”, the content of the target mention, and the IS prediction token “[IS]”. The previous overlap_info part contains two tokens, which indicate whether the target mention has the same string/head with a mention from the preceding sentences. And the local context is the sentence containing the target mention.

The final prediction is made based on the hidden state of the prediction token “[IS]”. In principle, the structure of the pseudo sentence and the mechanism of multi-head self-attention help the model to learn the important cues from both the mention and its discourse context when predicting IS.

3.3 Model Parameters

Our context-aware self-attention model has 12 transformer blocks, 768 hidden units, and 12 self-attention heads. We first initialize our model using BERTBASE, then fine-tune the model for 3 epochs with the learning rate of $5e-5$. During training and testing, the max token size of the pseudo sentence is set as 128.

4 Experiments

4.1 Experimental Setup

We perform experiments on the ISNotes corpus Markert et al. (2012), which contains 10,980 mentions annotated for information status in 50 news texts. Table 2 shows the IS distribution in ISNotes.

Following Hou et al. (2013a), all experiments are performed via 10-fold cross-validation on documents. We report overall accuracy as well as precision, recall and F-measure per IS class. In the following, we describe the baselines as well as our model with different settings.

collective (baseline1).

Hou et al. (2013a) applied collective classification to account for the linguistic relations among IS categories. They explored a wide range of features (34 in total), including a large number of lexico-semantic features (for recognizing bridging) as well as a couple of surface features and syntactic features.

cascaded collective (baseline2).

This is the cascading minority preference system for bridging anaphora recognition from Hou et al. (2013a).

self-attention wo context.

We apply our model (see Section 3) on the pseudo sentences containing only the target mentions and the prediction token “[IS]”.

self-attention with context I.

Based on self-attention wo context, we add the local context in the pseudo sentences.

self-attention with context II.

Based on self-attention with context I, we add the previous overlap_info part in the pseudo sentences.

4.2 Results and Discussion

Table 3 shows the results of our models compared to the baselines. Surprisingly, our model considering only the content of mentions (self-attention wo context) achieves competitive results as the baseline cascade collective which explores many hand-crafted linguistic features. Also self-attention wo context outperforms the two baselines on several IS categories (m/syntactic, m/aggregate, m/comparative, m/bridging and new). In Section 3.1, we analyze that m/syntactic and m/aggregate are often signaled by mentions’ internal syntactic structures, and that the semantics of certain premodifiers is a strong signal for m/comparative. The improvements on these categories show that our model can capture the semantic/syntactic properties of a mention when predicting its IS.

The continuous improvements on self-attention with context I and self-attention with context II show the impact of the local context and the previous context on IS prediction, respectively. It seems that the local context has more impact on m/bridging and new, whereas the previous context has more impact on old and m/worldKnowledge.

For self-attention with context I, we also tried to add the previous $k$ sentences ($k=1$ and $k=2$) into the current local context to see whether the broader local context can help us to capture bridging better. However, we found that the overall results on both settings are similar as the current one.

Overall, we achieve the new state-of-the-art results on bridging anaphora recognition with the local context model (self-attention with context I). And our full model (self-attention with context II) achieves an overall accuracy of 83% on IS classification, obtaining a 4.1% and 4.4% absolute improvements over the two baselines (collective and cascade collective), respectively. Our full model beats the two strong baselines on most IS categories except m/function. This is because there are only 65 m/function mentions in ISNotes. With such a small amount of training data, it is hard for our model to learn patterns for this category.

5 Conclusions

We develop a discourse context-aware self-attention model for IS classification. Our model does not contain any complex hand-crafted semantic features and achieves the new state-of-the-art results for IS classification and bridging anaphora recognition on ISNotes.