Natural Language Processing for
Human-Computer Interaction
Hae-Chang Rim
Korea University
Contents
• Introduction
• Conversational Natural Language Interface
• Language Understanding Components
• Dialog Management Models
• Conclusion
2
Introduction
1.
2.
3.
4.
What is Natural Language Processing (NLP)?
Two motivations for NLP
Research fields of NLP
NLP and HCI
Introduction
• What is Natural Language Processing (NLP)?
– This is a difficult question to answer since “there are almost as many
definitions as there are researchers studying it” (Obermeier, 1988)
The branch of information science that deals with natural language
information
The formulation and investigation of computationally effective mechanisms
for communication through natural language
A subfield of artificial intelligence and linguistics for making computers
"understand" statements written in human languages
• Two motivations for NLP (Allen, 1994)
– The scientific or linguistic motivation is to understand the nature of
language through the tools provided by computer science
– The technological motivation is to improve communication between
humans and machines
4
Introduction
• Research fields of NLP
Sentence Level
Morphological Analysis
Syntactic Analysis
Semantic Analysis
Speech Act Recognition
Reference Resolution
Planning & Reasoning
Language Generation
Question Answering
Conversational Agent
Discourse Level
Discourse Structure
Analysis
Dialog Level
Dialog Management
Application Level
Information Retrieval
And many other things
…
5
Introduction
• NLP and Human Computer Interaction (HCI)
– Goals of HCI (Bill 98)
• Developing systems which match or augment the physical,
perceptual, and cognitive capabilities of users
• Investigating a way to ensure the user-friendliness and robustness
of interactive computer systems
– NLP for HCI
• Since natural language is the most effortless and effective way of
communication in human-human interaction – either spoken or
typewritten, it may effectively complement other available
modalities
• Sometimes, natural language may even be the only applicable
modality:
– When driving car, carrying a baggage …
• NLP offers mechanisms for incorporating natural language
knowledge and modalities into user interfaces (Bill 98)
6
Conversational Natural Language
Interface
1. What is conversational NL interface system?
2. Limitation of current NLP techniques
3. Typical architecture of dialog system
Conversational NL interface
• What is conversational natural language interface system
(i.e. Dialog system)?
– Systems providing an interface that permits interaction through
natural language between the user and a computer-based
application
Hi, I’d like to fly to Seattle Tuesday morning.
Ok. Let’s see, I have a United flights ..
That’s OK
Will you return to Pittsburgh from …
8
Conversational NL interface
• Some limitations of current NLP techniques
– Full natural language understanding by machine may not be
realized in the near future
• Difficulty of resolving ambiguity of natural language
• Lack of resources …
• Practical Dialog Hypothesis (Allen et al. 01)
– Because of the limitation, current dialog systems have usually
been developed in a specific domain and for a specific task under
the practical dialog hypothesis
– Hypothesis:
• “The conversational competence required for practical dialogues,
while still complex, is significantly simpler to achieve than general
human conversational competence”
9
Conversational NL interface
• Typical Architecture of Dialog System
Language
Understanding
Components
Speech
Recognition
Dialog
Management
Component
QA Agents
Task Agents
Other Agents
Natural
Language
Generation
10
Conversational NL interface
• Two important issues related to NLP in building a dialog
system
Language
Understanding
Components
How to understand user’s utterance?
How to manage dialog between user and system?
Dialog
Management
Component
11
Language Understanding
Components
1. Overview of language understanding process
2. Morphological analysis
3. Part of speech (POS) tagging
3. Syntactic parsing
4. Semantic analysis
5. Discourse analysis
Language Understanding
Components
• The aim of language understanding components
– Analyze user’s utterance with discourse context and transform it
to semantic structure which the machine can understand
– Korean language understanding process
Morphological Analysis
POS Tagging
Syntactic Analysis
Semantic Analysis
Discourse Analysis
13
Finding all possible morphological
structure of a word (or Eojeol)
Disambiguate morphological
ambiguities
Finding a syntactic structure of an input
sentence
Finding a semantic structure without
using discourse context
Resolving remained ambiguity of
semantic analysis with discourse
context information
Morphological Analysis
• Problem domain
– Finding out
• Potential parts-of-speech for a given word (for English)
• Or morphological parses for a given Eojeol (for Korean)
– Morphological analyzer should produce all the grammatically
possible interpretations for a given word (or Eojeol)
– Example of morphological analysis in Korean
• When a sentence “나는 학교에 간다(na-neun hag-gyo-e
given, the result of morphological analysis is:
14
gan da)”
is
Morphological Analysis
• Difficulties of Korean morphological analysis
– Korean is
• A highly agglutinative languages
– An Eojeol is composed of one or more combined morphemes
• Very productive
– The number of Eojeols appeared in real texts is almost infinite
• A morphologically complex language
– Korean words (or Eojeols) are formed through compounding and
derivation
– Also morphological changes are frequently observed
»
“날/Nal/verb”+”는/Neun/connective_ending”  “나는/NaNeun”
• Hard to find the boundary of an unknown word
– In English, words (spacing units) which are not found in a dictionary
are unknown words
– In Korean, only subparts of them or themselves are unknown
morphemes
15
POS Tagging
• Part of Speech (POS) tagging is
– A task to assign a proper POS tag to each linguistic unit such as
word (in English), or morpheme (in Korean) for a given sentence
• An input of POS tagger is a result of morphological analysis, and
an output is a correct sequence of morpheme-POS pairs
– Hidden Markov Model (HMM) based POS Tagging
• Most popular and well-performed approach
– Regard POS tags of morphemes in a given sentence as hidden states
and find the most probable path in a lattice
16
Syntactic Parsing
• Goal
– Find out a syntactic structure with a specific grammar for a given
sentence
• Example of parsed sentence with the phrasal structure grammar
– “누나는 예쁜 꽃을 좋아한다. (Nu-Na-Neun ye-Ppeun Kkoch-eul Coh-a-HanTa.)”
S
VP
VP
NP
NP
Nu-Na Neun
NC
JX
17
ADJP
Ye-Ppeu n
PA
EM
NP
Kkoch eul
NC
JC
Coh-a-Han-Ta
PV+EF
.
SS.
Syntactic Parsing
• Parsing can be defined as
– A problem that maps any input sentence to an appropriate
syntactic tree structure (Chung, 04)
– Why is the parsing so difficult?
• Because of the structural ambiguity of natural language!
• Several characteristics of Korean make the parsing more difficult
– Relatively free-word order, constituent ellipsis …
S
S
VP
VP
NP
18
어제
유진이
VP
쇼를
보았
다
보았다
어제
유진이
쇼를
보았다
Syntactic Parsing
• Examples of statistical parsing with simple PCFG model
– “Astronomers saw stars with ears”
TreeBank
S NP VP
VP V NP
VP VP PP
PP  P NP
P  with
V  saw
NP  NP PP
0.4
NP  astronomers 0.1
NP  ears
0.18
NP  saw
0.04
NP  stars
0.18
NP  telescope
0.1
1.0
0.7
0.3
1.0
1.0
1.0
P(t1)
=
0.0009072
t1
19
P(t2)
=
0.0006804
>
t2
Semantic Analysis
• Semantic analysis is
– The process whereby meaning representations are composed and
assigned to a user’s utterance
How can I go to Korea
University?
20
What does
it mean?
Semantic Analysis
• Example of semantic analysis
21
Semantic Analysis
• Shallow semantic analysis for a dialog system
– Under the practical dialog hypothesis, we can simplify the
semantic analyzing process:
• Restricting domain of a dialog system reduce the ambiguity
– In the pay-bill domain, the word `bank’ may not be used as the
meaning of a dike
• Also, if we restrict a task of a dialog system, simple methods such
as concept-spotting can be enough to capture user’s intention
“6시에 MBC에서 뭐 하니?”
Analyzed by a concept spotting method
Question Focus: Program
Channel: MBC
Begin_time: 18:00
22
Discourse Analysis
• Reference resolution
– The omitted words (or phrases) and the pronominal references
are complemented by the use of common sense and discourse
information
U: I would like to open a fixed
deposit account.
S: For what amount?
U: Make it for 8000 dollars.
• Speech Act Identification
– Speech Act: The communicative intention represented by each
utterance
– A dialog system should have the ability to
• identify other participants’ speech act, predict next possible speech
acts, and generate own utterance suitable for the speech act
23
Statement
non
opinion:
Statement
opinion:
I'm a customer since November.
I think it's great.
How to manage dialog between user
and system?
1. Overview of dialog management
2. FST based Approach
3. Frame based Approach
4. Other Approaches
Dialog Management
• Dialog management model (or component)
– Controlling the flow of the dialog between the system and the
user, including the coordination of other components of the
system
• Dialog management model must solve two problems:
– Keep track of the overall interaction with steady progress
towards task completion
• The system must have some idea of the task completion ratio
• More importantly, the system must have some idea of what is yet to
be done,
– Robustly handle deviations from the nominal progression
towards problem solution
25
Dialog Management
• One of core issues of dialog management
– System-initiative:
• system always has control, user only responds to system questions
– User-initiative:
• user always has control, system passively answers user questions
– Mixed-initiative:
• control switches between system and user
• Classification of dialog management strategies (Allen et
al. 01)
–
–
–
–
26
Finite state (or graph)-based strategy
Frame-based strategy
Plan-based strategy
Agent-based strategy
Dialog Management
• Finite-state based dialog control
– Simplest dialog control method
– Usually, system-initiative
– Dialogue consists of a sequence of predetermined steps or states
• The dialog flow is specified as a set of dialogue states with
transitions denoting various alternative paths through the dialog
graph
• Most commercially available spoken dialog system use this form of
dialogue management strategy
– Example task: Long distance dialing by voice, Tele-banking system
– Does not require sophisticated NLP techniques, but works only
for simple tasks
27
Dialog Management
Example of finite-state based dialog management: “Pay a bill”
28
Dialog Management
• Example illustrating some limitations of finite-state
based dialog system
The over-informative answer cannot be accepted
I already answered for that question!
29
Dialog Management
• Frame-based dialog management
– More flexible approach
– Mixed Initiative using fixed rules
– Dialog management problem is regarded as form filing :
• The form specifies all relevant information (slots) for an action
– Dialog management consist of
• Monitoring the form for completion
• Extract relevant elements from user utterance
• Asking question to user using empty slots as a trigger
30
Dialog Management
• Example of Frame-based Dialogue Control
Frame:Send a message
Importance
of Slots
Filling a slot by a user response
Triggering a system response by an empty slot
31
Dialog Management
• More complex dialog management approaches
– Plan (Task) Based Model: The dialogue involves interactively
constructing a plan (e.g. kitchen design consultant).
– Agent Based Model: Involves planning and also executing and
monitoring operations in a dynamically changing world (e.g.
emergency rescue coordination).
– Generally require deep semantic analysis for user utterances,
rich knowledge resources, and elaborate inference/reasoning
methods
32
Dialog Management
• Summary of dialog management approaches
33
Features/Dialog
control strategies
Finite state-based
Frame-based
Plan/Agent-based
Input
Single words or phrases
Natural language with
concept spotting
Unrestricted natural language
Verification
Explicit confirmation
Explicit and implicit
confirmation
Grounding
Dialogue model
Information state represented
implicitly in dialog states
Dialog control represented
explicitly with state diagram
Explicit representation of
information states
Dialog control represented
with control algorithm
Model of system’s intention,
goals, and beliefs
Dialog history, context
Dialogue
phenomena
User answers question
User asks question, simple
clarifications by system
Dynamically generated topic
structure, collaborative
negotiation subdialgues
Different modalities (e.g.,
planned and actual world
Example Task
Long-distance dialing
Getting train arrival and
departure information
Kitchen design consultant,
Disaster relief management
Conclusion
• Conversational NL interface are showing promise as a
new modality for HCI, because
– Natural language is most familiar way of communication in
human-human interaction
– It also can provide “effortless and effective” way of
communication in a human-computer interaction
• However, there are still serious obstacles to be overcome
– Improving performances of NLP analysis components such as
POS tagging, parsing, so on
– Ensure domain portability of a dialog interface-based system
– …
34
Descargar

학위논문발표