Asian American Pacific Islander Month DDSD 2024.pptx
Lecture 1: Semantic Analysis in Language Technology
1. Semantic Analysis in Language Technology
Lecture 1: Introduction
Course Website: http://stp.lingfil.uu.se/~santinim/sais/sais_fall2013.htm
MARINA SANTINI
PROGRAM: COMPUTATIONAL LINGUISTICS AND LANGUAGE TECHNOLOGY
DEPT OF LINGUISTICS AND PHILOLOGY
UPPSALA UNIVERSITY, SWEDEN
12 NOV 2013
2. Acknowledgements
2
Thanks to Mats Dahllöf for the many slides I
borrowed from his previous course and for
structuring such an interesting and comprehensive
content.
Lecture 1: Introduction
5. Check the website regularly and make sure to refresh the page:
we are building up this course together, so this page will be continously
updated!
5
Lecture 1: Introduction
6. About the Course
6
Introduction to Semantics in Language Techology
and NLP.
Focus on methods used in Language Technology
and NLP for the perform the following tasks:
Sentiment Analysis (SA)
Information Extraction (IE)
Word Sense Disambiguation (WSD)
Predicate-Argument Extraction (PAS)
Lecture 1: Introduction
7. Intended Learning Outcomes
7
In order to pass the course, a student must be able to:
describe systems that perform the following tasks, apply them to authentic
linguistic data, and evaluate the results:
1.
detect and extract attitudes and opinions from text, i.e. Sentiment
Analysis (SA);
2.
use semantic analysis in the context of Information Extraction (IE)
3.
disambiguate instances of polysemous lemmas, i.e. Word Sense
Disambiguation (WSD);
4.
use robust methods to extract the Predicate-Argument Structure (PAS).
Lecture 1: Introduction
8. Compulsory Readings
8
1.
Bing Liu (2012) Sentiment Analysis and Opinion Mining, Morgan & Claypool.
2.
Richard Johansson and Pierre Nugues. 2008. Dependency-based Syntactic–
Semantic Analysis with PropBank and NomBank, CoNLL 2008: Proceedings
of the 12th Conference on Computational Natural Language Learning.
3.
Daniel Jurafsky and James H. Martin (2009), Speech and Language
Processing: An Introduction to Natural Language Processing, Computational
Linguistics, and Speech Recognition. Second Edition, Pearson Education.
4.
Daniel Gildea and Daniel Jurafsky. 2002. Automatic Labeling of Semantic
Roles, Computational Linguistics 28:3, 245-288.
5.
M Palmer, D Gildea, P Kingsbury. 2005. The proposition bank: An annotated
corpus of semantic roles, Computational Linguistics 31 (1), 71-106.
6.
Additional suggested readings will be listed at the end of each lecture
Lecture 1: Introduction
9. Demos & Tutorials
9
This list will be continuosly updated, also with your
contribution…
Lecture 1: Introduction
10. Assignments and Examination
10
Four Assignments:
1.
Essay writing: independent study of a system, an approach, or a field within
semantics-oriented language technology. The study will be presented both as a written
essay and an oral presentation. The essay work will also include a feedback step where
the work of another group is reviewed.
2.
Assignment on Predicate-Argument Structure (PAS)
3.
Assignment on Sentiment Analysis (SA)
4.
Assignment on Word Sense Disambiguation (WSD)
General Info:
No lab sessions, supervision by email
Essay and assignments must be submitted to santinim@stp.lingfil.uu.se
Examination:
Written report submitted for each assignment
All four assignments necessary to pass the course
Grade G will be given to students who pass each assignment. Grade VG to those who
pass the essay assignment and at least one of the other ones with distinction.
Lecture 1: Introduction
11. IMPORTANT!
11
Start thinking about a topic you are interested in for
your essay writing assignment!
Lecture 1: Introduction
12. Practical Organization
12
45min + 15 min break
Lectures on Course webpage and SlideShare
Email all your questions to me: santinim@stp.lingfil.uu.se
IMPORTANT:
Send me an email to santinim@stp.lingfil.uu.se, so I make sure that I have
all the correct email addresses. If you do not get an acknowledgement of
receipt, please give me a shout!
Lecture 1: Introduction
13. Interaction and Cooperation
13
Communicate with me and with your classmates to
exchange ideas, if you have problems in
understanding notions and concepts or practical
implementations.
Recommemdation: share your knowledge with your
peers and steam off stress.
Cheating is not permitted
Lecture 1: Introduction
14. Semantics in Language Technology Overview
14
SEMANTICS IN LANGUAGE TECHNOLOGY
APPLICATIONS
LEXICAL SEMANTICS
REPRESENTATION OF MEANING
SUMMARY
Lecture 1: Introduction
16. Logic and Semantics
16
Aristotelian logic – important ever since.
Syllogisms, e.g.:
Premise: No reptiles have fur.
Premise: All snakes are reptiles.
Conclusion: No snakes have fur.
Modern logic develops, late 19th Century – more
general and systematic.
Formal semantics in linguistics and philosophy
based on logic (20th Century).
Lecture 1: Introduction
17. Formal and Computational Semantics
17
Computational semantics “is the study of how to
automate the process of constructing and reasoning
with meaning representations of natural language
expressions.” (Wikipedia).
Early systems rule-based, most famous example:
“Montague grammar” (1970). Sophisticated
mechanisms for translation of English into a very
rich logic.
Language technology: Recent interest in data-driven
and machine learning-based methods.
Lecture 1: Introduction
18. Semantics in NLP
18
NLP semantics is typically more limited in scope than NL
semantics as analysed in linguistics and philosophy.
NLP applications often handle semantic aspects without
having explicitly semantic components, e.g. in machine
translation.
Other aspects of language – morphology, syntax, etc. –
can be seen as support systems for semantics: The
purpose of language lies in the use of expressions as
carriers of semantic meaning. And that is what many
NLP systems have to respect, e.g. MT, retrieval,
classification, etc.
Lecture 1: Introduction
19. Semantics and Truth (i)
19
Semantics, meanings and states of affairs:
What a sentence means: a structure involving
(lexical) concepts and relations among them.
Can be articulated as a semantic
representation.
E.g. I ate a turkey sandwich. in predicate logic:
A sentence and the semantic representation of
a sentence is also the representation of a
possible state of affairs.
Lecture 1: Introduction
20. Semantics and Truth (ii)
20
Correspondence theory of truth: If the content of a sentence
corresponds to an actual state of affairs if it is true; otherwise, it is
false.
Ignoring philosophical complications, in many cases we can extract
knowledge from texts.
E.g. Warmer climate entails increased release of carbon
dioxide by inland lakes. (From uu.se press release.)
Related issue: Which texts should we trust?
Many sentences are difficult to formalize in logic. (Modality,
conditionality, vague quantification, tense, etc.)
Lecture 1: Introduction
22. Formalizing Meaning
22
Linguistic content has – at least to a certain degree – a logical
structure that can be formalized by means of logical calculi –
meaning representations.
The representation languages should be simple and
unambiguous – in contrast to complex and ambiguous NL.
Logical calculi come with accounts of logical inference. They
are useful for reasoning-based applications.
Meaning formalization faces far-reaching conceptual and
computational difficulties.
Lecture 1: Introduction
23. Compositionality
23
Linguistic content is compositional: Simple
expressions have a given (lexical) meaning; the
meaning of complex expressions is determined by
the meanings of their constituents.
People produce and understand new phrases and
sentences all the time. (NLP must also deal with
these.)
Compositionality is studied in detail in
compositional syntax-driven semantics. Work in
this field is typically about hand-coded rule systems
for small fragments of NL.
Lecture 1: Introduction
28. First-Order Predicate Logic (i)
28
“flexible, well-understood, and computationally
tractable approach to the representation of
knowledge [and] meaning” (J&M. 2009: 589)
expressive
verifiability against a knowledge base (related to
database languages)
inference
model-theoretic semantics
Lecture 1: Introduction
29. First-Order Predicate Logic (ii)
29
Boolean operators: negation and connectives
Existential/universal quantification
Individual constants
Predicates (taking a number of arguments)
Lecture 1: Introduction
30. When to assume compositionality?
30
Lecture 1: Introduction
31. Multi-Word Expressions
31
MWEs (a.k.a multiword units or MUs) are lexical units
encompassing a wide range of linguistic phenomena,
such as idioms (e.g. kick the bucket = to die), collocations
(e.g. cream tea = a small meal eaten in Britain, with
small cakes and tea), regular compounds (cosmetic
surgery), graphically unstable compounds (e.g. selfcontained <> self contained <> selfcontained - all
graphical variants have huge number of hits in Google),
light verbs (e.g. do a revision vs. revise), lexical bundles
(e.g. in my opinion), etc. While easily mastered by
native speakers, MWEs' correct interpretation remains
challenging both for non-native speakers and for
language technology (LT), due to their complex and often
unpredictable nature.
Lecture 1: Introduction
32. Cross-linguality
Use Case: Information Access
32
In multi-ethnic societies, like the Swedish society, it is common that many non-native speakers
use public websites – e.g. Arbetesförmedlingen or Pensionsmyndigheten websites – to access
information that are vital to their living and integration in the host country. National
regulations are often accompanied by special terminology and new coinages. For instance, the
Swedish expression /egenremiss/ (14,900 hits, Google.se April 2013) – or alternatively as an
MWE – /egen remiss/ (8,210 hits, Google.se April 2013) denotes a referral to a specialist doctor
written by patients themselves. This expression is made up from two common Swedish words
/egen/ `own (adj)' and /remiss/ `referral'. It is a recent expression (probably coined around
20101) and not yet recorded in any official dictionary nor in Wiktionary or other multilingual
online lexical resources. However, it is very frequent in query logs belonging to a Swedish public
health service website. When trying to implement a cross-lingual search based on the automatic
translation of query logs, it turned out that none of the existing multilingual lexical resources
contained this expression.
Lecture 1: Introduction
33. Use Case: Personal Use & Text Understanding
33
The use of expressions that are marked for style, genre, domain, or
register (and/or other textual categories), or the use of expressions
which are misspelled or idiomatic for some textual category are
beyond the competence of a novice reader or a non-native speaker.
Additionally, in a web search or in social networks, one cannot tell if
the texts one reads are good or bad the way a firstlanguage readers
can. When readers/users read a language they do not know at all,
they can use automatic translation or online dictionaries or other
lexical resources. However, what they cannot determine well is the
*type* of text one is reading. They cannot tell if the text is
verbose, terse, formal, informal, stupid, funny, bad, or good.
For instance, the phrase "es ist zum Kotzen" means this is
vernacular and unrefined text as well as a controversial expression.
The phrase "isch alle", instead, means that this line in the text is
spoken by a Berliner.
Lecture 1: Introduction
34. Semantics vs Pragmatics/Discourse (i)
34
What does a word, a phrase, a text segment mean as an
NL expression? (“Linguistic meaning” – semantics.)
Conventional, static, systemic aspect of meaning.
What does the author intend to convey by means of a
word, a phrase, a text segment? (“Speaker meaning” –
pragmatics/discourse.)
Contextual, dynamic aspect of meaning.
The two aspects depend on each other, of course.
Lecture 1: Introduction
38. Semantics-oriented NLP applications
38
Machine translation: The translation of a text segment should
mean the same as the original (to emphasize linguistic
meaning) or should convey the same content (to emphasize
speaker meaning).
Information extraction is to extract components of the
information conveyed by a text.
Question answering is extraction – combined with inference –
of an answer to a given question.
Text classification, in typical cases, relates to the meanings of
the texts being classified.
Lecture 1: Introduction
39. Semantics and Generation
39
Generation: semantic representation NL. Less
challenging than analysis – the structure of the input
is under control. Needed in e.g. dialogue systems.
Interlingua – semantic representation in machine
translation:
Analysis: source language interlingua.
Generation: interlingua target language.
Would be economic if many languages are involved. The idea has
not proved very successful so far.
Lecture 1: Introduction
40. Reference
40
Reference is very important – what statements are
about.
Referring expressions are very common.
Reference is a discourse phenomenon.
Resolving reference is a crucial step in e.g.
extraction, e.g.in sentiment analysis
translation, e.g. to get agreement right
English it vs French il/elle vs Swedish den/det.
Lecture 1: Introduction
42. Kinds of Referring Expressions
42
Indefinite noun phrases. E.g. a book. Introduce new
entities.
Pronouns. E.g. he. Typically coreferent with a
previous referring expression (antecedent).
Names. E.g. Bill Gates.
Demonstrative. E.g. this room.
Other definite noun phrases. E.g. the first chapter.
Reference to somehow known entity, often
previously mentioned.
Lecture 1: Introduction
43. Named Entity Recognition (NER)
43
To identify expressions being used as names. (What
characterizes a “name”?)
Also to identify what kind of name it is: E.g. of a
person, or a place, or a stretch of time, or a chemical
compound, or a gene, etc.
“State-of-the-art NER systems for English produce
near-human performance. For example, the best
system entering MUC-7 scored 93.39% of F-measure
while human annotators scored 97.60% and 96.95%”
(Wikipedia).
Lecture 1: Introduction
44. Anaphora and Deixis Resolution
44
Pronouns (they), pronominal adverbs (there, then), and
definite NP’s refer to entities by means of contextually
given information.
E.g. by referring to previously mentioned referents –
anaphora.
E.g. by reference based on the participants, time, and
place of the discourse – deixis (e.g. I, you, here,
yesterday).
Anaphora and deixis resolution is much more
challenging task than NER. The reference of name-like
graph words is much more predictible. Compare Barack
Obama and he.
Lecture 1: Introduction
45. Sentiment Analysis – an extraction task
45
What views do people express in blogs and reviews?
That’s interesting for politicans and marketing people.
Opinions are often expressed in a personal and informal
way.
E.g. Peter bought me a Baileys marzipan chocolate thing
which I washed down with Gluehwein and that, in
combination with the bright lights and cheery faces really
made me feel warm inside! (From a blog post.)
Sentiment analysis: to extract the referent of a
“sentiment” and the polarity positive–negative
associated with it.
E.g. Baileys marzipan chocolate – positive.
Lecture 1: Introduction
47. Lexical Concepts
47
Words are often grammatically simple, but carry a
structured conceptual content. Definitions “unpack”
the content of concepts:
friend – a person whom one knows well, is loyal to, etc.
turkey – a kind of animal, a bird, etc.
sandwich – a kind of food item, contains bread , etc.
eat – a relation (holding in/of an event) between an organism
and a food item, the food is chewed and ingested, etc.
Lecture 1: Introduction
51. Synonimy
51
Synonymy holds between two words (word tokens) which express the same
or similar concepts.
Unsupervised detection of synonymy can be based on “The Distributional
Hypothesis: words with similar distributions have similar meanings.” =
The Distributional Hypothesis in linguistics is the theory that words
that occur in the same contexts tend to have similar meanings. The
underlying idea that "a word is characterized by the company it
keeps" was popularized by Firth.
“Random Indexing” is a method here. (“a high-dimensional model can be
projected into a space of lower dimensionality without compromising
distance metrics if the resulting dimensions are chosen appropriately”)
Synonymy knowledge useful in e.g. translation, text classification, and
information extraction. Also “query expansion” in retrieval.
Lecture 1: Introduction
56. Word Senses
56
Discerning word senses (for a lemma) –
lexicographical task, matter of sophisticated
linguistic judgements.
Theoretical principles. Practical purpose.
Different dictionaries make different analyses.
English: WordNet – a standard resource.
Lecture 1: Introduction
57. Senses of day in WordNet, for instance (i)
57
Lecture 1: Introduction
58. Senses of day in WordNet, for instance (ii)
58
Lecture 1: Introduction
59. Word Sense Disambiguation (WSD)
59
A distributional hypothesis for WSD: words representing
the same sense have more similar distributions than
words representing different senses.
I.e. distribution similarity implies sense similiarity.
We can use this for supervised learning of WSD.
This requires data in the form of a sense-tagged corpus
(based on a given sense inventory, e.g. the one given by
WordNet).
Lecture 1: Introduction
60. Manual Sense-Tagging
60
More difficult than typical grammatical tagging.
As we saw in the day example, senses and their
distinctions can be quite subtle. Definitions and
examples are often far from obvious.
Expensive: requires competent people and standardised
procedures.
Quality measure: inter-annotator agreement. ” Ex:
Cohen's kappa coefficient is a statistical measure
of inter-rater agreement or inter-annotator
agreementfor qualitative (categorical) items. It is
generally thought to be a more robust measure than
simple percent agreement calculation since κ takes into
account the agreement occurring by chance ”
Lecture 1: Introduction
62. Conclusions (i)
62
Logic-based semantics is a theoretical foundation for
NLP semantics, but implemented systems are
typically more coarse-grained and of a more limited
scope.
Meaning depends both on literal content and
contextual information. This is a challenge for most
NLP tasks.
Most NLP applications have to be highly sensitive to
semantics.
Lecture 1: Introduction
63. Conclusions (ii)
63
Finding and interpreting names and other referential
expressions is a central issue for NLP semantics.
Disambiguation of polysemous lexical tokens is also
a central issue for NLP semantics.
Accessing the content of lexical tokens is also useful.
Meaning representation involves predicateargument structure, which captures a basic aspect of
NL compositionality.
Lecture 1: Introduction
64. 64
Start thinking about a Topic of interest for
your essay writing! Tell me your thoughts
next time…
Lecture 1: Introduction
65. Suggested Readings
65
Term Logic (Wikipedia)
Predicate Logic (Wikipedia)
Jurafsky and Martin (2009):
Ch. 17 ”Representation of Meaning”
Ch. 18 ”Computational Semantics”
Ch. 19 ”Lexical Semantics”
Ch. 20 ”Compuational Lexical Semantics”
Clark et al. (2010):
Ch 15 ”Computational Semantics”
Indurkhya and Damerau (2010):
Ch 5 ”Semantic Analysis”
Lecture 1: Introduction
66. 66
This is the end… Thanks for your attention !
Lecture 1: Introduction