Mentions légales du service

Skip to content
Snippets Groups Projects

Compare revisions

Changes are shown as if the source revision was being merged into the target revision. Learn more about comparing revisions.

Source

Select target project
No results found
Select Git revision

Target

Select target project
  • mastertal/discourse-parsing
  • chuyli/discourse-parsing
2 results
Select Git revision
Show changes
Commits on Source (12)
Hide whitespace changes
Inline Side-by-side
Showing
with 0 additions and 1179 deletions
discourse-parsing.html deleted 100644 → 0
View file @ f1f1fd94
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8" />
<meta name="keywords" content="remark,remarkjs,markdown,slideshow,presentation" />
<meta name="description" content="A simple, in-browser, markdown-driven slideshow tool." />
<title>DiscourseParsing</title>
<style>
/* modified to point to our local separate files */
@import url("common/fonts.css");
@import url("common/style.css");
</style>
</head>
<body>
<textarea id="source">
class: center, middle
background-image:url(images/data-background-light.jpg)
# Discourse parsing
<br>
### Discourse analysis - Discourse corpora - Discourse parsers
https://gitlab.com/cbraud/coursediscourse
<br><br>
#### Nancy, 16 January 2020
<br><br>
.pull-left[<img src="images/logo-ul.png" width="50%"/>]
.pull-right[.bold[Chloé Braud, [chloe.braud@irit.fr](mailto:chloe.braud@irit.fr),
CNRS-IRIT]]
<br><br>
---
class: middle
# Discourse analysis
*Whenever we read something closely, with even a bit of sensitivity, text
structure leaps off the page at us. We begin to see elaborations, explanations,
parallelisms, contrasts, temporal sequencing, and so on. These relations bind
contiguous segments of text into a global structure for the text as a whole.*
(Hobbs, 1985)
---
# Discourse processing
### What is discourse?
* Document-level analysis
* Coherence and cohesion
* Discourse analysis
* Theoretical frameworks
### Discourse parsing
* Discourse corpora
* Discourse parsing and discourse chunking
* Applications
* Current challenge
### Practical session
* Explicit vs implicit discourse relations
---
.left-column[
## What is discourse?
### - Document-level
]
.right-column[
### Multi sentence linguistic phenomena
* Topics (topic segmentation),
* Temporal links,
* Entities and reference,
* Rhetorical/discourse relations
Discourse structure is about:
* revealing text coherence,
* interpreting documents (i.e. making inferences on its content),
There are links between the different kind of text organization, e.g.:
* constraints discourse/temporal
* e.g. often the effect after the cause
* discourse/topic
* e.g. some relations require to keep the same topic
* discourse/coreference
* e.g. some relations block a potential referent
]
---
.left-column[
## What is discourse?
### - Document-level
]
.right-column[
### Multi sentence linguistic phenomena
* Topics (topic segmentation),
* Temporal links,
* Entities and reference,
* .alert[Rhetorical/discourse relations]
Discourse structure is about:
* revealing text coherence,
* interpreting documents (i.e. making inferences on its content),
There are links between the different kind of text organization, e.g.:
* constraints discourse/temporal
* e.g. often the effect after the cause
* discourse/topic
* e.g. some relations require to keep the same topic
* discourse/coreference
* e.g. some relations block a potential referent
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
]
.right-column[
## Coherence and cohesion
Document: not a random sequence of sentences
* A text is **cohesive** if its elements are linked together (non structural
textual relations)
* A text is **coherent** if it makes sense (structural relation between segments).
$\rightarrow$ Document = coherent structured group of sentences
Simple examples:
.small[
* *Paul fell. Marie helped him get up.* (Narration)
* *Paul fell. Marie pushed him.* (Cause)
* *Paul fell. He likes spinach.* (??)
* *Paul went to Istanbul. He likes travelling.* (Explanation)
* *Paul went to Istanbul. He likes spinach.* (? Explanation)
]
People tend to "force" a meaningful relation between sentences!
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
]
.right-column[
#### Cohesive but not coherent text:
i.e.: Each sentence is notionally linked to the one that precedes it, using both
lexical and grammatical means, but the text is ultimately senseless
.small[*I am a teacher. The teacher was late for class. Class rhymes with grass. The
grass is always greener on the other side of the fence. But it wasn't.* (Teacher
resource site)]
#### Automatically generated summary: not cohesive, not coherent
i.e.: improper sentence ordering, pronoun without antecedent
.small[*It’s like going to disney world for car buyers. I have to say that Carmax rocks.
We bought .alert[it] at Carmax, and I continue to have nothing bad to say about
that company. After our last big car milestone, we’ve had an odyssey with cars.*
[Mithun and Kosseim, 2011]]
#### The task of modeling coherence / cohesion:
* "discern an original text from a permuted ordering of its sentences"
* "locate the original position of a sentence previously removed "
* "compare the rankings, given by the model, against human judgments"
* goal: e.g., coherence scoring (automatically evaluating student essays,
L2 learners productions, ...), or improving text generation, summarization
]
---
background-image: url(images/conn-lex.jpg)
background-size: 380px
background-position: 5% 90%
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
]
.right-column[
### Discourse Relations
The existence of discourse relations is hinted by discourse connectives, such
as *however*, *moreover*, *meanwhile*, *if..then*...
These connectives:
* contribute to cohesion and coherence
* explicitly specify the relation between adjacent units of text:
* *however* signals a contrastive relation
* *moreover* signals that the subsequent text elaborates or strengthens
the point that was made immediately beforehand,
* *meanwhile* indicates that two events are contemporaneous
* *if...then* sets up a conditional relationship.
**Connective lexicons:** [connective-lex](http://connective-lex.info/)
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
]
.right-column[
### Discourse Relations link:
* the semantic contents of two units (1)
* or the speech act expressed and the semantic content (2)
* same phenomena within sentences (3)
Examples:
.small[
* This cute child turns out to be a blessing and a curse.
*She gives the Artist a sense of purpose, but also alerts him to the serious
inadequacy of his vagrant life.* **(Cause-reason)**
]
.small[
* Mrs. Yeargin is lying. *They found students (..) who said she gave them
similar help.* **(Pragmatic Cause-justification)**
]
.small[
* Typically, money-fund yields beat comparable short-term investments
*because portfolio managers can vary maturities and go after highest rates.*
**(Cause-reason)**
]
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
]
.right-column[
### Partial consensus:
* Analysis unit: a document
* Elementary Discourse Unit (EDU): mostly clauses, at most a sentence
* Discourse relations: semantico-pragmatic, binary, inter- and intra-sentential
* Explicit: with a discourse connective
* Implicit: without a discourse connective
Examples:
.small[
* The tawny owl is a nocturnal bird of prey, .alert[**but**]
it can live in the daytime.
]
.small[
* The towers collapsed less than two hours later .alert[**(Result)**] dragging
down with them the building of the Marriott World Trade Center.
.alert[**(Sequence)**] The tower 7 of the WTC collapsed in the afternoon
.alert[**because**] of damages caused by the fall of Twin Towers.
]
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
### - Theoretical frameworks
]
.right-column[
### Frameworks/annotation schemes: 2 views
#### Hierarchical discourse structure (RST, SDRT, DLTAG, GraphBank...)
* Structure: trees/graphs covering the documents
* Try to give an interpretation to the document
* Annotation is hard!
#### Local coherence (PDTB)
* "theory neutral": lexically grounded
* Flat structure, no full covering
* Higher inter-annotator agreement (larger corpora)
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
### - Theoretical frameworks
]
.right-column[
### Different frameworks/annotation schemes
#### Rhetorical Structure Theory DT [carlson et al., 2001]
* Relations: definition based on the author intentions
* 78 relations, 16 classes
* One relation per pair of units
* Structure: trees covering the documents
#### Penn Discourse TreeBank (PDTB) [Prasad et al., 2008]
* Annotation based on connectives and adjacency
* Hierarchy: 4 classes, 16 types, 23 subtypes
* Possibly multiple relations
* Flat/no structure: no relation between some units
#### SDRT: Annodis [Afantenos et al. 2012]
* Relation definitions based on semantics
* 18 relations
* One relation per pair of units, embedded units
* Structure: graphs covering the documents
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
### - Theoretical frameworks
]
.right-column[
#### RST
<img src="images/rst3.png" width="75%"/>
.west[
#### SDRT
<img src="images/sdrt-salmon.png" width="95%"/>
]
.east[
#### PDTB
<img src="images/pdtb.jpg" width="60%"/>
]
]
---
.left-column[
## What is discourse?
### - Document-level
### - Coherence
### - Discourse analysis
### - Theoretical frameworks
]
.right-column[
*Strong generative capacity of RST, SDRT and discourse dependency DAGSs*,
Danlos, Constraints in Discourse, 2008
<img src="images/comparison.png" width="65%"/>
]
---
# Discourse parsing
<img src="images/rst42-4.png" width="20%"/>
<img src="images/rst42-3.png" width="70%"/>
<img src="images/rst42-2.png" width="100%"/>
<img src="images/rst42-1.png" width="100%"/>
---
.left-column[
## Discourse parsing
### - Corpora
##### RST DT
]
.right-column[
### The RST Discourse TreeBank
[RST website](https://www.sfu.ca/rst/)
* Annotated upon the PTB
* 385 documents
* RST analysis goal: recovering the author’s "intentions"
* Typically one “more important” segment (nucleus vs satellite)
* Most used for discourse parsing because:
* **Trees!**: similar to syntactic parsing
* full coverage of documents: all parts are connected
Honnestly: old, weird corpus...
* only 1 relation per pair of segments
* very strange relations: *attribution*, *same-unit*...
* very strange segmentation: see below
* the definition of the relations is very hard to understand:
[take a look](https://www.sfu.ca/rst/01intro/definitions.html)
.small[*|Mr. Volk, 55 years old, succeeds Duncan Dwight,| |who retired in September.|*]
.small[*|The Tass news agency said the 1990 budget anticipates income of 429.9 billion rubles| |($US693.4 billion)| *]
]
---
.left-column[
## Discourse parsing
### - Corpora
##### RST DT
]
.right-column[
### Relation set (classes and relations)
* **Attribution**: attribution, attribution-negative
* **Background**: background, circumstance
* **Cause**: cause, result, consequence
* **Comparison**: comparison, preference, analogy, proportion
* **Condition**: condition, hypothetical, contingency, otherwise
* **Contrast**: contrast, concession, antithesis
* **Elaboration**: elaboration-additional, elaboration-general-specific, elaboration-part-whole,
elaboration-process-step, elaboration-object-attribute, elaboration-set-member, example, definition
* **Enablement**: purpose, enablement
* **Evaluation**: evaluation, interpretation, conclusion, comment
* **Explanation**: evidence, explanation-argumentative, reason
* **Joint**: list, disjunction
* **Manner-Means**: manner, means
* **Topic-Comment**: problem-solution, question-answer, statement-response, topic-comment,
comment-topic, rhetorical-question
* **Summary**: summary, restatement
* **Temporal**: temporal-before, temporal-after, temporal-same-time, sequence, invertedsequence
* **Topic Change**: topic-shift, topic-drift
]
---
.left-column[
## Discourse parsing
### - Corpora
##### PDTB
]
.right-column[
### The Penn Discourse Treebank
[PDTB Website](https://www.seas.upenn.edu/~pdtb/)
* Wall Street Journal Articles
* Annotated upon the Penn Treebank
* 2,259 documents
#### Annotation
* Explicit relations: 18,459
* Connectives: closed-list of 100
* Arguments = minimal text spans
* Implicit relations: 16,224
* Alternative lexicalizations: 624
* Entity Relations: 5,210
* No Relation: 254
* **40,600** annotations
]
---
.left-column[
## Discourse parsing
### - Corpora
##### PDTB
]
.right-column[
### Explicit discourse relations
* Annotation: the connective is annotated, the relation(s) triggered (up to 2)
and the arguments
* Different types of connectives (and positions):
* Subordinating conjunctions (e.g., because, when, since, although)
.small[
|Use of dispersants was approved|1 **when** |a test on the third day showed some positive
results|2, officials said. (CONTINGENCY:Cause:reason)
**Although** |the purchasing managers’ index continues to indicate a slowing economy,|2
|it isn’t signaling an imminent recession|1, said Robert Bretz (COMPARISON:Concession:expectation)
]
* Coordinating conjunctions (e.g., and, or, nor):
.small[
The theory is |that Seymour is the chief designer of the Cray-3,| **and**
|without him it could not be completed.|2
(EXPANSION.Conjunction and CONTINGENCY.Cause.result)
]
* (ADVP and PP) adverbials (e.g., however, otherwise, then, as a result, for example)
.small[
A Chemical spokeswoman said |the second-quarter charge was "not material"|1 |and
that no personnel changes were made|2 **as a result**. (CONTINGENCY:Cause:result)
]
]
---
.left-column[
## Discourse parsing
### - Corpora
##### PDTB
]
.right-column[
### Implicit discourse relations
.small[
|Mrs Yeargin is lying.|1 **Implicit = because** |They found students in an advanced class
a year earlier who said she gave them similar help.|2 (CONTINGENCY:Pragmatic
Cause:justification)
]
* Annotated between adjacent segments (only sentences in PDTB 2)
* Up to 4 relations annotated per pair of segments
* Rule: a connective can be inserted
* and the connective is annotated
* Introduce a bias?
* Adding a connective may hide a relation
* Some papers: try to guess this connective and then the relation
* Require this specific annotation to be extended to other languages, genres...
]
---
.left-column[
## Discourse parsing
### - Corpora
##### PDTB
]
.right-column[
### Alternative lexicalizations
* "cases where a discourse relation is inferred between adjacent sentences"
* "but where providing an Implicit connective leads to redundancy in the
expression of the relation"
* "the relation is alternatively lexicalized by some “non-connective expression”
.small[
And she further stunned her listeners by revealing her secret garden design
method: |Commissioning a friend to spend “five or six thousand dollars . . .
on books that I ultimately cut up.”|1 **AltLex After that**,
|the layout had been easy.|2
The Bank of England, on the other hand, had gold reserves that averaged about
30% of its outstanding currency (...)
**AltLex The most likely reason for this disparity** is that (...)
]
]
---
.left-column[
## Discourse parsing
### - Corpora
##### PDTB
]
.right-column[
### Relation set
<img src="images/pdtb-rel.png" width="100%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
]
.right-column[
* RST: many new projects for other languages, often simplifying the
relation set
* PDTB: many languages covered (but rarely full annotation, eg. French = only
connectives)
<img src="images/corpora.png" width="100%"/>
]
---
background-image: url(images/steps-nlp.jpg)
background-size: 200px
background-repeat: no-repeat
background-position: center
background-size: 100%
class:inverse
## Discourse parsing
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
]
.right-column[
### Natural Language Processing:
* many applications just focus on sentence processing (eg. see the problems
previously shown for summarization)
* often try to at least take co-reference into account
* discourse information is "needed" (if needed) at the end of the pipeline
* also means that discourse processing needs all the information from the
previous steps (you can easily imagine the problem with error propagation ...)
* one thing that makes discourse parsing hard
Next step: Pragmatic Analysis
* *In this step, data is interpreted on what it actually meant. Although, we have
to derive aspects of language which require real-world knowledge.*
* reference to objects in the world,
* but also deal with context, implicature, etc
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
]
.right-column[
### Discourse parsing: first step, segmenting
* Segment a document into EDUs
* mostly clauses and sentences, but a bit more fine-grained in the RST DT
* see the large set of rules in the [tagging manual](https://www.isi.edu/~marcu/discourse/tagging-ref-manual.pdf), eg:
.small[
* Includes both speech acts and other cognitive acts:
|The company says| |it will shut down its plant.|
* But if the complement is a to-infinitival, do not segment:
|The company wants to shut down its plant.|
* But segment infinitive clause marking a purpose relation (but not all of them,
would be too easy...):
|A grand jury has been investigating
whether officials (...) conspired **to** cover up their accounting| |**to**
evade federal income taxes.|
]
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
]
.right-column[
### Discourse segmenters
Most existing systems use lexical, POS tags and syntactic information + gold
sentence segmentation (not a so easy task!)
* RST DT: [Xuan Bach et al.]: F1 91.0% (automatic parse) / 93.7% (gold parse)
* English instructional corpus: [Joty et al, 2015] F1 80.9%
### ToNy, winner of the last shared task :)
See the results [here](https://sites.google.com/view/disrpt2019/shared-task?authuser=0),
and the paper [here](https://hal.archives-ouvertes.fr/hal-02374091/file/21_Paper.pdf)
[Muller et al, 2019]
* Using contextual embeddings alone allows close to state-of-the-art results
* ELMo better than BERT on English, but not multilingual
* Results with BERT multilingual, average over the languages:
* F1 90.11% if sentence boundaries given
* F1 86.38% else
* Problem with cross-domain learning:
* Training on GUM and testing on RST-DT: drop from 96% to 66%
* Training on RST-DT to test on GUM from 93% to 73%
* Note: [GUM corpus](http://corpling.uis.georgetown.edu/gum/) is composed of
documents from several domains.
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
]
.right-column[
### Discourse parsing: second step, building the tree
* Attachment: which EDUs are linked together
* Labeling: with which relation / sense
* Recursive process: the pair of discourse units has to be linked to another
segment, and so on, until full coverage
* + RST bonus: label each segment as nucleus or satellite
Parsers are inspired from syntactic parsing:
* Transition based, shift-reduce, CKY parsing (constituency or dependency)
* Main problems:
* Efficiency: trees are often far deeper than in syntax
* Representation: we need to encode spans of text instead of just words
* Relations are semantic, harder to identify than syntactic ones
* Lack of data: corpora are small, 385 documents in the RST DT, meaning
385 trees / instances for ou system
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
]
.right-column[
### Representing discourse units and their combination [Ji and Eisenstein, 2014](https://www.aclweb.org/anthology/P14-1002.pdf)
* Idea: jointly learn the task and the word representation (as low dimensional
vector)
* Test 3 options for transforming the original features, taking into account
relationships between adjacent EDUs
### Overcoming the lack of data by splitting the task [Wang et al, 2017](https://www.aclweb.org/anthology/P17-2029.pdf)
* Idea: not enough data for structure + nuclearity + relation.
* First: build a parser that identifies the naked structure + nuclearity
* Then: relations, 3 classifiers (within/across sentence, across paragraphs)
[Morey et al, 2017](https://hal.archives-ouvertes.fr/hal-01650251/document): evaluation problem, scores in [Ji and Eisenstein, 2014] are
not computed using the right evaluation metrics, F1=57.8% (and not 61.6%)
<img src="images/rst-parsing.png" width="40%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
]
.right-column[
### What about other languages? [Braud et al., 2017](https://arxiv.org/pdf/1701.02946.pdf)
* Cross-lingual experiments:
* Train only on data for other languages
* Train on data for other languages but optimize the hyper-parameters on data
for the target language
* Transfer is very hard!
* Monolingual experiments: large drop of performance for language other than
english, ie. smaller corpora
<img src="images/rst-cross.png" width="100%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Discourse chunking or shallow discourse parsing
* Identifying discourse connectives
* Identifying connective arguments:
* position
* boundaries
* Identifying the sense of the discourse relation (label)
<img src="images/pdtb-pipeline2.png" width="80%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Interesting subtasks
#### Explicit discourse relations: Connective ambiguity
* Usage ambiguity:
* discourse reading vs non discourse reading: whether or not a given token
is serving as a discourse connective in its context
* e.g. in 1. *since*: no discourse reading
* Sense ambiguity:
* what discourse relation(s) a given token is signalling
* e.g. in 2. *since* signals a temporal relation while in 3 *since* signals
a cause
.small[
1. She has been up **since** 5am.
2. There have been over 100 mergers **since** the most recent wave of friendly takeovers ended.
3. It was a far safer deal **since** the company has a healthier cash flow
]
- [Pitler and Nenkova, 2009]: syntactic features are useful for both tasks
- [Webber et al. 2019]: more on connective ambiguity
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Interesting subtasks
#### Identifying discourse connectives (usage)
* Closed-list of known discourse connectives:
* *because, in the meanwhile, but, if..then, on the one hand..on the other hand...*
* Disambiguation problem:
* *Paul likes dogs **and** cats* (no discourse reading)
* *Paul feeds the cat **and** pets the dog.* (discourse reading)
* Binary classification
* Syntactic and lexical features: connective, its POS-tag and
immediate context, syntactic sisters and path to root
* **High performance: around 95% in F1**
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Interesting subtasks
#### Identifying connective arguments
* Relative position of Arg1 and Arg2:
* Classification: same/previous(/following) sentence
* Representation: position of the connective, context
* High performance: 97.94% in F1
* Exact span:
* Select the nodes in the syntactic tree included
* **Moderate performance: 53.85-86-24% in F1**
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Interesting subtasks
#### Explicit sense classifier
* Multiclass classification
* Features: connective, connective postag, previous and following postag and
word
* **High performance: 86.77% in F1**
But, be careful, that's the general picture:
* performance drops when tokenization or pos tagging are not manual / gold
* performance drops for other languages or domains
* good performance for higly frequent connectives, or a few very unambiguous
connectives
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
See [Johannsen and Sogaard, 2013]
<img src="images/conn-scores-sogaard.png" width="65%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Interesting subtasks
#### Implicit sense classifier
* Multiclass classification
* Features: word pairs, modality, semantic classes, polarity...
* **Low performance: 42-57.1% in F1** (level 1-2 relations)
* a very hard task (let's try during the practical!), yet not solved
* but crucial: about 50% of the relations are implicit
Many strategies tested:
* Semi-supervision/domain adaptation:
* using explicit examples, automatically annotated data
* Distant supervision:
* building a word representation tailored to the task
* multi-task learning using data from other discourse corpora, or data for
other tasks (temporality, co-reference, speech acts, ...)
* Various algorithms:
* including all variations of neural networks
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Explicit vs implicit relations
**Results of the last shared task**: compare F-measure for Explicit vs Implicit
<img src="images/results_explimpl.png" width="110%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
]
.right-column[
### Full pipeline on the PDTB (except AltLex, EntRel, Norel, Attribution)
**Results of the last shared task**: max 27.7 in F-measure!!
<img src="images/results_conll_2016.png" width="100%"/>
<img src="images/results_conll_2016_chinese.png" width="100%"/>
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
### - Applications
]
.right-column[
### Discourse structure: useful for several tasks and applications
* Temporal ordering
* Co-reference [Cristea et al., 1999]
* Automatic summarization [Sporleder and Lapata, 2005]
* Question Answering [Verberne, 2007]
* Sentiment analysis [Bhatia et al., 2015]
* Essay scoring [Higgins et al., 2004; Mesgar and Strube, 2018]
* Summary coherence rating [Nguyen and Joty, 2017], coherence Modeling
[Li and Jurafsky, 2017; Mesgar and Strube, 2018]
* Readability assessment
* Machine translation [Meyer and Webber, 2013; Born et al., 2017 ]
.small[
*Paul fell, Mary pushed him.* Explanation $\rightarrow$ pushed < fell
*The champions league has become a source of income for clubs **since** it
started in 1992.* Temporal
*La ligue des champions est devenue une source de revenus pour les clubs
**car** il a commencé en 1992.* Causal
]
]
---
.left-column[
## Discourse parsing
### - Corpora
### - Discourse processing
### - RST parsing
### - PDTB parsing
### - Applications
### - Current challenges
]
.right-column[
### Room for improvement everywhere, but more importantly:
* Evaluation methods:
* stop evaluating only on RST DT or PDTB (i.e. English + Wall Street Journal)
* be careful when evaluating, see [Morey et al. 2017]
* use downstream applications: is all this work really useful? Do we really
need (full) discourse parsers? We need to see that it helps for other tasks
* Learning from a limited amount of data (especially with neural methods):
* we can't expect unlimited amount of annotated data, especially for discourse
* Adapting to new domains or languages:
* even languages and domains without any annotated data [Braud et al. 2017]
* Conversations / Dialogues: only a few corpora + annotation very hard
* automatically summarizing meetings
* improving chatbots
* detecting schizophrenia using patient-doctor conversations ;)
* useful for social media analysis: Twitter / forum are similar to dialogues
]
---
count: false
## Sources and reading list
.small[
* Course on discourse parsing at ESSLLI 2019: https://github.com/TScheffler/2019ESSLLI-discparsing
* J. eisenstein's NLP course: https://github.com/jacobeisenstein/gt-nlp-class/blob/master/notes/eisenstein-nlp-notes-10-15-2018.pdf
* Old slides: http://www.dfki.de/~horacek/09-Discourse-Parsing.pdf
* ConLL shared task 2015: http://www.cs.brandeis.edu/~clp/conll15st/
* ConLL shared task 2016: http://www.cs.brandeis.edu/~clp/conll16st/
* C. Pott's course on discourse: http://compprag.christopherpotts.net/pdtb.html
* ICDM 18 tutorial: https://drive.google.com/file/d/1XmaN6tXxnVasw8Sp0cr0FXbG7KHsEJlM/view
and https://drive.google.com/file/d/1QcbkKGZI8BAZh3v0_36BKDHw-rpOLjVC/view
* Discourse corpora:
* RST tagging manual: https://www.isi.edu/~marcu/discourse/tagging-ref-manual.pdf
* Discourse processing:
* Disambiguating explicit discourse connectives without oracles, Johannsen and Sogaard, IJCNLP, 2013
https://www.aclweb.org/anthology/I13-1134.pdf
* Applications:
* Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation,
Läubli et al., EMNLP, 2018, https://www.aclweb.org/anthology/D18-1512.pdf
* Modeling local coherence: An entity-based approach, Barzilez and Lapata, ACL 2005,
https://people.csail.mit.edu/regina/my_papers/coherence.pdf
* Automatically Evaluating Text Coherence Using Discourse Relations, Lin et all, ACL, 2011
* A Neural Local Coherence Model, Nguyen and Joty, ACL 2017,
https://www.aclweb.org/anthology/P17-1121.pdf
* More references on coherence modeling in ICDM 18 slides
* Steps in NLP (picture): https://data-flair.training/blogs/ai-natural-language-processing/
]
<section data-background-iframe="https://nbviewer.jupyter.org/urls/mastertal.gitlab.io/UE803/notebooks/Exercise_sheet_7.ipynb" data-background-interactive>
</section>
</textarea>
<script src="common/remark-latest.min.js"></script>
<script>
var hljs = remark.highlighter.engine;
</script>
<script src="common/remark.language.js"></script>
<script src="common/mermaid/mermaid.min.js"></script>
<script src="common/katex/katex.min.js"></script>
<script src="common/katex/contrib/auto-render.min.js"></script>
<script src="common/terminal.language.js" type="text/javascript"></script>
<link rel="stylesheet" href="common/mermaid/mermaid.css">
<link rel="stylesheet" href="common/katex/katex.min.css">
<script>
var options = {
highlightStyle: 'monokai',
highlightLanguage: 'remark',
highlightLines: true,
// Set the slideshow display ratio
// Default: '4:3'
// Alternatives: '16:9', ...
ratio: '16:9',
};
var renderMath = function() {
//renderMathInElement(document.body);
// or if you want to use $...$ for math,
renderMathInElement(document.body, {delimiters: [ // mind the order of delimiters(!?)
{left: "$$", right: "$$", display: true},
{left: "$", right: "$", display: false},
{left: "\\[", right: "\\]", display: true},
{left: "\\(", right: "\\)", display: false},
]});
}
var slideshow = remark.create(options, renderMath) ;
// don't let mermaid automatically load on start
mermaid.initialize({
startOnLoad: false,
cloneCssStyles: false
});
function initMermaidInSlide(slide) {
var slideIndex = slide.getSlideIndex();
// caution: no API to get the DOM element of current slide in remark,
// this might break in the future
var currentSlideElement = document.querySelectorAll(".remark-slides-area .remark-slide")[slideIndex];
var currentSlideMermaids = currentSlideElement.querySelectorAll(".mermaid");
if (currentSlideMermaids.length !== 0) {
mermaid.init(undefined, currentSlideMermaids);
}
}
// first starting slide won't trigger the slide event, manually
// init mermaid
initMermaidInSlide(slideshow.getSlides()[slideshow.getCurrentSlideIndex()]);
// on each slide event, trigger init mermaid
slideshow.on('afterShowSlide', initMermaidInSlide);
// extract the embedded styling from ansi spans
var highlighted = document.querySelectorAll("code.terminal span.hljs-ansi");
Array.prototype.forEach.call(highlighted, function(next) {
next.insertAdjacentHTML("beforebegin", next.textContent);
next.parentNode.removeChild(next);
});
</script>
</body>
</html>
DiscourseParsing.pdf → discourse-parsing_part1.pdf
View file @ 9b22411c
No preview for this file type
discourse-parsing_part2.pdf 0 → 100644
View file @ 9b22411c
File added
getSlides.sh deleted 100755 → 0
View file @ f1f1fd94
#! /bin/bash
# contact: yannick.parmentier@loria.fr
# date: 2019/07/08
name=$1
if [ $# -lt 1 ];
then
echo "Usage :"
echo " $0 <file_name>"
exit 1
else
cp .slides-template.html $name
echo "Slides file created."
fi
exit 0
images/Imprimer la commande n°143819976.pdf deleted 100644 → 0
View file @ f1f1fd94
File deleted
images/UZWSUE_1501_BRAUD_1a1f03fc60102.pdf deleted 100644 → 0
View file @ f1f1fd94
File deleted
images/applications.png deleted 100644 → 0
View file @ f1f1fd94
images/applications.png

77.6 KiB

images/argmicrotext.png deleted 100644 → 0
View file @ f1f1fd94
images/argmicrotext.png

117 KiB

images/comparison.png deleted 100644 → 0
View file @ f1f1fd94
images/comparison.png

64.1 KiB

images/comparison2.png deleted 100644 → 0
View file @ f1f1fd94
images/comparison2.png

81.6 KiB

images/conn-lex.jpg deleted 100644 → 0
View file @ f1f1fd94
images/conn-lex.jpg

281 KiB

images/conn-scores-sogaard.png deleted 100644 → 0
View file @ f1f1fd94
images/conn-scores-sogaard.png

95.5 KiB

images/connectives.png deleted 100644 → 0
View file @ f1f1fd94
images/connectives.png

121 KiB

images/corpora.png deleted 100644 → 0
View file @ f1f1fd94
images/corpora.png

89.3 KiB

images/cross-seg.png deleted 100644 → 0
View file @ f1f1fd94
images/cross-seg.png

81.2 KiB

images/data-background-light.jpg deleted 100644 → 0
View file @ f1f1fd94
images/data-background-light.jpg

221 KiB

images/disrpt.jpeg deleted 100644 → 0
View file @ f1f1fd94
images/disrpt.jpeg

5.95 KiB

images/gears3.jpg deleted 100644 → 0
View file @ f1f1fd94
images/gears3.jpg

74.9 KiB

images/logo-cnrs.png deleted 100644 → 0
View file @ f1f1fd94
images/logo-cnrs.png

143 KiB

images/logo-ul.png deleted 100644 → 0
View file @ f1f1fd94
images/logo-ul.png

63.8 KiB