The Intersection of Robust Intelligence and Trust in Autonomous Systems: Papers from the AAAI Spring Symposium
[THIS SPACE MUST BE
KEPT BLANK]
Improving Trust in Automation of Social Promotion
Boris Galitsky, Dmitri Ilvovsky, Nina Lebedeva and Daniel Usikov
Knowledge-Trail Inc. San Jose CA 95127 USA
bgalitsky@hotmail.com; dilv_ru@yahoo.com; nina.lebedeva_kt@gmail.com; usikov@hotmail.com
Abstract
domain where users are more tolerant to agent’s
misunderstanding of what they said.
In this study we build a conversational agent enabled with
robust intelligence for social promotion. It is presented as a
simulated human character which acts on behalf of its
human host to facilitate and manage her communication
for her. The agent relieves its human host from the routine,
less important activities on social networks such as sharing
news and commenting on messages, blogs, forums, images
and videos of others. Unlike the majority of application
domains for simulated human characters, its social partners
do not necessarily know that they exchange news,
opinions, and updates with an automated agent. We refer to
this agent as a conversational agent which assists its
human host with social promotion (CASP). We will
experiment with CASP in a number of Facebook accounts
and evaluate its performance and trust by human users
communicating with it.
To be trusted, a conversational character operating in a
natural language must produce the content relevant to what
the human peers share. To do that, it needs to implement
robust intelligence features (NSF 2013, Lawless et al
2013):
1. Flexibility in respect to various forms of human
behavior, and human mental states, such as
information sharing, emotion sharing, and
information request by humans.
2. Resourcefulness, being capable of finding
relevant content in emergent and uncertain
situations, for a broad range of emotional and
knowledge states.
3. Creativity in finding content and adjusting
existing content to the needs of human peers.
4. Real-time responsiveness and long-term reflection
on how its postings being perceived. CASP builds
a taxonomy of entities it used for relevance
assessment and reuses it in future sessions
(Galitsky 2013).
5. Use of a variety of reasoning approaches, in
particular based on simulation of human mental
states.
We build a conversational agent performing social
promotion (CASP) to assist in automation of interacting
with friends and managing other social network contacts.
This agent employs a domain-independent natural language
relevance technique which filters web mining results to
support a conversation with friends and other network
members. This technique relies on learning parse trees and
parse thickets (sets of parse trees) of paragraphs of text such
as Facebook postings. We evaluate the robust intelligent
features and trust in behavior of this conversational agent in
the real world. It is confirmed that maintaining a relevance
of automated posting is essential in retaining trust and
efficient social promotion of CASP.
Introduction
Simulated human characters and agents are increasingly
common components of user interfaces of applications
such as interactive tutoring environments, eldercare
systems, virtual reality-based systems, intelligent assistant
systems, physician-patient communication training, and
entertainment applications (Cassell et al., 2000, De Rosis
et al., 2003; Dias and Paiva 2005; Lisetti 2008, among
others). While these systems have improved in their
intelligent features, expressiveness, understanding human
language, dialog abilities and other aspects, their social
realism is still far behind. It has been shown (Reeves and
Nass 1996) that users consistently respond to computers as
if they were social actors; however most systems don’t
behave as competent social actors, leading to user loose of
trust and alienation. Most users used to distrust
conversational agent who has shown poor understanding of
their needs in the areas such as shopping, finance, travel,
navigation, customer support and conflict resolution. To
restore trust in automated agents, they have to demonstrate
robust intelligence features on one hand and operate in a
Copyright © 2014, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
28
6.
Ability to learn and adapt performance at a level
of intelligence seen in humans and animals.
7. Awareness of and competence in larger natural,
built, and social contexts.
For a conversational system, users need to feel that it
properly reacts to their actions, and that what it replied
makes sense. To achieve this in a limited, vertical domain,
most effective approaches rely on domain-specific
ontologies. In a horizontal domain, one needs to leverage
linguistic information to a full degree to be able to
exchange messages in a meaningful manner. Once we do
not limit the domain a conversational agent is performing
in, the only available information is language syntax and
discourse, which should be taken into account in a full
scale linguistic relevance filter (Galitsky et al 2012).
Social promotion is based on
• involvement (living the social web, understanding it,
going beyond creation of Google+ account);
• creating (making relevant content for communities of
interest);
• discussing (each piece of content must be accompanied
by a discussion. If an agent creates the content the market
needs and have given it away freely, then you
will also want to be available to facilitate the
ensuing discussions);
• promoting (the agent needs to actively,
respectfully, promote the content into the
networks).
CASP acts in the environments subject to
constant changes. As news come, political
events happen, new technologies are developed
and new discoveries are made, CASP needs to
be able to find relevant information using new
terms or new meanings of familiar terms and
entities. Hence it needs to automatically learn
from the web, expanding its taxonomy of
entities and building links between them
(Galitsky 2013). These taxonomies are
essential when CASP needs to match a portion
of text found on the web (as a candidate
message) against a message posted by a human
user. By growing these taxonomies, CASP
learns from the web, adapts its messages to how the
available information on the web is evolving. Also, CASP
applies accumulated experience from user responses to its
previously posted messages to new posting new problems
may be better suited to respond to novel environmental
challenges..
Paragraphs of text as queries appear in the search-based
recommendation domains (Montaner et al., 2003; Bhasker
and Srikumar 2010; Thorsten et al., 2012) and social
search (Trias et al., 2012). Recommendation agents track
user chats, user postings on blogs and forums, user
comments on shopping sites, and suggest web documents
and their snippets, relevant to a purchase decisions. To do
that, these recommendation agents need to take portions of
text, produce a search engine query, run it against a search
engine API such as Bing or Yahoo, and filter out the search
results which are determined to be irrelevant to a purchase
decision. The last step is critical for a sensible functionality
of a recommendation agent, and poor relevance would lead
to lost trust in the recommendation engine. Hence an
accurate assessment of similarity between two portions of
text is critical to a successful use of recommendation
agents.
Parse trees have become a standard form of representing
the syntactic structures of sentences (Abney 1991; de Salvo
Braz et al 2005; Domingos and Poon 2009). In this study
we represent a linguistic structure of a paragraph of text
based on parse trees for each sentence of this paragraph.
We will refer to the set of parse trees plus a number of arcs
for inter-sentence relations between nodes for words as
Parse Thicket. A PT is a graph which includes parse trees
for each sentence (Collins and Duffy 2002), as well as
additional arcs for inter-sentence relationship between
parse tree nodes for words.
Fig. 1: Social promotion facilitated by CASP
We will rely on the operation of generalization of text
paragraphs via generalization of respective parse thickets
to assess similarity between them.
The domain of social promotion
On average, people have 200-300 friends or contacts on
social network systems such Facebook and LinkedIn. To
maintain active relationships with this high number of
friends, a few hours per week is required to read what they
29
post and comment on it. In reality, people only maintain
relationship with 10-20 most close friends, family and
colleagues, and the rest of friends are being communicated
with very rarely. These not so close friends feel that the
social network relationship has been abandoned.
However, maintaining active relationships with all
members of social network is beneficial for many aspects
of life, from work-related to personal. Users of social
network are expected to show to their friends that they are
interested in them, care about them, and therefore react to
events in their lives, responding to messages posted by
them. Hence the users of social network need to devote a
significant amount of time to maintain relationships on
social networks, but frequently do not possess the time to
do it. For close friends and family, users would still
socialize manually. For the rest of the network, they would
use CASP for social promotion being proposed.
The difficulty in solving this problem lies mainly in the
area of relevance. A robust intelligence paradigm is
expected to come into play to solve this problem in real
world environment with different cultures, broad range of
user interests and various forms of friendship and
professional connections in social networks.
Messages of the automated agent must be relevant to
what human agents are saying. These messages are not
always expected to be impressive, witty, or written in style,
but at least they should show social engagement, it should
show that the host of CASP cares about the friend being
communicated with.
The opportunity to automate social promotion leverages
the fact that overall coherence and exactness of social
communication is rather low. Readers would tolerate worse
than ideal style, discourse and quality of content being
communicated, as long as overall the communication is
positive and makes sense. Currently available commercial
chat bots employed by customer support portals, or
packaged as mobile apps, possess too limited NLP, text
understanding and overall robust intelligent capabilities to
support conversations to support social promotion.
For the purpose of promoting social activity and enhance
communications with the friends other than most close
ones, CASP is authorized to comment on postings, images,
videos, and other media. Given one or more sentence of
user posting or image caption, this agent issues a Bing
Web and Bing Blogs APIs search request and filters the
results for relevance.
answer to a recommendation post seeking more questions,
etc.
CASP architecture
CASP includes the following components (Fig. 3):
• Web mining component, which forms the web search
queries from the seed.
• Content relevance component, which filters out
irrelevant portions of candidate content found on the web,
based on syntactic generalization operator. It functions
matching the parse forest for a seed with the parse forest
for a content found on the web.
• Mental state relevance component, which extracts
Fig. 2: CASP is about to post a message about his “experience” at
lake Tahoe, having his host’s friend image caption as a seed.
Web mining for the content relevant to the seed:
1) Forming a set of web search queries
2) Running web search and storing candidate
portions of text
Content relevance verification:
Filtering out candidate postings with low parse thicket
generalization scores
Mental states relevance verification:
Filtering out candidate postings which don’t form a
plausible sequence of mental states with the seed
Fig.3: Three main components of CASP.
CASP inputs a seed (a posting written by a human) and
outputs a message it forms from a content mined on the
web and adjusted to be relevant to the input posting. This
relevance is based on the appropriateness in terms of
content and appropriateness in terms of mental state: for
example, it responds by a question to a question, by an
30
mental states from the seed message and from the web
content and applies reasoning to verify that the former can
be logically followed by the latter.
mental states and behaviors of human agents in broad
domains where training sets are limited.
The system architecture serves as a basis of OpenNLP –
similarity component, which is a separate Apache Software
foundation project, accepting input from either OpenNLP
or Stanford NLP. It converts parse thicket into JGraphT
objects which can be further processed by an extensive set
of graph algorithms (Ehrlich and Rarey, 2011). The code
and
libraries
described
are
available
at
http://code.google.com/p/relevance-based-on-parse-trees
http://svn.apache.org/repos/asf/opennlp/sandbox/opennlpsimilarity/. The web mining and posting search system is
pluggable into Lucene library to improve search relevance.
Also, a SOLR request handler is provided so that search
engineers can switch to a parse thicket-based multisentence search to quickly verify if relevance is improved
in their domains.
Building Parse Thicket
Obtain parse trees for
all sentences
Find communicative actions and
set relations between them and
their subjects
Obtain coreferences
relations between nodes
Find rhetoric markers and
form rhetoric structure
relations
Generalizing phrases for two Parse Thickets
Obtain all phrases occurring in
one sentence
Obtain thickets phrases by
merging ones from distinct
sentences
Relevance assessment operation
Group phrases by type
(NP, VP, PP)
For two sets of thicket phrases, select
a pair of phrases. Align them and find
a maximal common sub-phrase
Remove sub-phrases from the list
of resultant phrases
Modern search engines are not very good at tackling
queries including multiple sentences. They either find a
very similar document, if it is available, or very dissimilar
ones, so that search results are not very useful to the user.
This is due to the fact that for multi-sentences queries it is
rather hard to learn ranking based on user clicks, since the
number of longer queries is practically unlimited. Hence
we need a linguistic technology which would rank
candidate answers based on structural similarity between
the question and the answer. In this study we build a graphbased representation for a paragraph of text so that we can
track the structural difference between these paragraphs,
taking into account not only parse trees but the discourse
as well.
To measure the similarity of abstract entities that are
expressed by logic formulas, a least-general generalization,
or anti-unification, was proposed for a number of ma-chine
learning approaches, including explanation-based learning
and inductive logic programming. We extend the notion of
generalization from logic formulas to the sets of syntactic
parse trees of these portions of text. Rather than extracting
common keywords, the generalization operation produces
a syntactic expression that can be semantically interpreted
as a common meaning shared by two sentences.
In this section we briefly describe a contribution to
Apache OpenNLP Similarity, which takes the constituency
parse trees and finds a set of maximal common sub-trees as
expressions in the form of a list {<lemma, part-of-speech>,
…}.
Let us represent a meaning of two NL expressions by
using logic formulas and then construct the generalization
(^) of these formulas: camera with digital zoom and
camera with zoom for beginners
Generalize Speech
Act & Rhetoric
relation
Calculate the score for
maximal common subphrases
score
Fig. 4: Generalizing parse thicket to assess relevance of the seed
text and candidate mined portion of text (the second component
in Fig. 3).
In (Galitsky 2013) we developed a generic software
component for computing consecutive plausible mental
states of human agents, which is an important part of a
robust intelligence and is employed by CASP. The
simulation approach to reasoning about mental world is
based on exhaustive search through the space of available
behaviors. This approach to reasoning is implemented as a
logic program in a natural language multiagent mental
simulator NL_MAMS, which yields the totality of possible
mental states few steps in advance, given an arbitrary
initial mental state of participating agents. Due to an
extensive vocabulary of formally represented mental
attitudes, communicative actions and accumulated library
of behaviors, NL_MAMS is capable of yielding much
richer set of sequences of mental state than a conventional
system of reasoning about beliefs, desires and intentions
would deliver. Also, NL_MAMS functions in domainindependent manner, outperforming machine learningbased systems for accessing plausibility of a sequence of
31
To express the meanings, we use the predicates
camera(name_of_feature,
type_of_users)
and
zoom(type_of_zoom). The above NL expressions will be
represented as follows: camera(zoom(digital), AnyUser) ^
camera(zoom(AnyZoom), beginner), where the variables
(uninstantiated values that are not specified in NL
expressions) are capitalized. The generalization is
camera(zoom(AnyZoom), AnyUser). At the syntactic level,
we have the generalization of two noun phrases as follows:
{NN-camera, PRP-with, NN-zoom]}.
We use (and contribute to) the (OpenNLP 2013) system
to derive the constituent trees for generalization (chunker
and parser). The generalization operation occurs on the
levels of Paragraph/Sentence/Phrases (noun, verb and
others)/Individual word. At each level, except for the
lowest level (individual words), the result of the
generalization of two expressions is a set of expressions. In
this set, expressions for which less-general expressions
exist are eliminated. The generalization of two sets of
expressions is a set of the sets that are the results of the
pair-wise generalization of the expressions in the original
two sets.
Algorithm for generalizing two constituency parse trees
Input: a pair of sentences
Output: a set of maximal common sub-trees
Obtain the parsing tree for each sentence, using OpenNLP.
For each word (tree node), we have a lemma, a part of
speech and the form of the word’s information. This
information is contained in the node label. We also have an
arc to the other node.
1) Split sentences into sub-trees that are phrases for
each type: verb, noun, prepositional and others.
These sub-trees are overlapping. The sub-trees are
coded so that the information about their occurrence
in the full tree is retained.
2) All of the sub-trees are grouped by phrase types.
3) Extend the list of phrases by adding equivalence
transformations (Galitsky et al 2012).
4) Generalize each pair of sub-trees for both sentences
for each phrase type.
5) For each pair of sub-trees, yield an alignment and
generalize each node for this alignment. Calculate the
score for the obtained set of trees (generalization
results).
6) For each pair of sub-trees of phrases, select the set of
generalizations with the highest score (the least
general).
7) Form the sets of generalizations for each phrase type
whose elements are the sets of generalizations for
that type.
8) Filter the list of generalization results: for the list of
generalizations for each phrase type, exclude more
general elements from the lists of generalization for a
given pair of phrases.
Evaluation of content relevance
Having described the system architecture, we proceed to
evaluation of how generalization of PTs can improve
multi-sentence search, where one needs to compare a query
as a paragraph of text against a candidate posting as a
paragraph of text (snippet). We conducted evaluation of
relevance of syntactic generalization – enabled search
engine, based on Yahoo and Bing search engine APIs. For
an individual query, the relevance was estimated as a
percentage of correct hits among the first thirty, using the
values: {correct, marginally correct, incorrect}. Accuracy
of a single search session is calculated as the percentage of
correct search results plus half of the percentage of
marginally correct search results. Accuracy of a particular
search setting (query type and search engine type) is
calculated, averaging through 40 search sessions. For our
evaluation we use user postings available at author’
Facebook accounts. The evaluation was conducted by the
authors. We refer the reader to (Galitsky et al 2012) for the
further details on evaluation settings.
Fig. 5: Assessing similarity between two sentences.
The similarity between two sentences is expressed as a
maximum common subtree between their respective parse
trees (Fig. 5). The algorithm that we present in this paper
concerns the paths of syntactic trees rather than subtrees
because these paths are tightly connected with language
phrases (Galitsky et al 2013).
32
Complexity of the seed and
posted message
A friend complaints to the
CASP’s host
A friend unfriends the CASP
host
A friend shares with other
friends that the trust in CASP is
lost
A friend encourages other
friends to unfriend a friend with
CASP
6.3
8.5
9.4
12.8
2 sent
6.0
8.9
9.9
11.4
Relevance of parse thicket-based graph
generalization search, %, averaging over
100 searches
1 sent
Relevance of thicket-based phrase
generalization search, %, averaging over
100 searches
Facebook
friend
agent
support
search
Topic of
the seed
Relevance single-sentence phrase-based
generalization search, %, averaging over
100 searches
Query
complexity
Relevance of baseline Bing search, %,
averaging over 100 searches
Query
type
1compound
sent
2 sent
74.5
83.2
85.3
87.2
72.3
80.9
82.2
83.9
3 sent
5.9
7.4
10.0
10.8
3 sent
69.7
77
81.5
81.9
4 sent
5.2
6.8
9.4
10.8
4 sent
70.9
78.3
82.3
82.7
1 sent
7.2
8.4
9.9
13.1
2 sent
6.8
8.7
9.4
12.4
3 sent
6.0
8.4
10.2
11.6
4 sent
5.5
7.8
9.1
11.9
1 sent
7.3
9.5
10.3
13.8
2 sent
8.1
10.2
10.0
13.9
3 sent
8.4
9.8
10.8
13.7
4 sent
8.7
10.0
11.0
13.8
1sent
3.6
4.2
6.1
6.0
2 sent
3.5
3.9
5.8
6.2
3 sent
3.7
4.0
6.0
6.4
4 sent
3.2
3.9
5.8
6.2
1 sent
7.1
7.9
8.4
9.0
2 sent
6.9
7.4
9.0
9.5
3 sent
5.3
7.6
9.4
9.3
4 sent
5.9
6.7
7.5
8.9
Travel &
outdoor
Shopping
Table 1: Content relevance evaluation results
To compare the relevance values between search settings,
we used first 30 search results and re-ranked them
according to the score of the given search setting. We use
three approaches to verify relevance between the seed text
and candidate posting:
1) Pair-wise parse tree matching, where the tree for
each sentence from seed is matched with the tree
for each sentence in candidate posting mined on
the web;
2) The whole graph (parse thicket) for the former is
matched against a parse thicket for the latter using
phrase-based approach. In this case parse thickets
are represented by all paths (linguistically,
phrases)
3) The match is the same as 2) but instead of phrases
we find a maximal common subgraph (Galitsky et
al 2013).
One can observe that unfiltered precision is 58.2%,
whereas improvement by pair-wise sentence generalization
is 11%, thicket phrases – additional 6%, and graphs –
additional 0.5%. One can also see that the higher the
complexity of sentence, the higher the contribution of
generalization technology, from sentence level to thicket
phrases to graphs.
Events &
entertain
ment
Jobrelated
Personal
life
Average
6.03
7.5
8.87
10.575
Table 2: Evaluation results for trust loosing scenarios.
Trust in the CASP can be lost in the following scenarios:
1) The content CASP is posted is irrelevant to the
content of original post by a human friend.
2) It is relevant to the content, but irrelevant in terms
of mental state.
3) Both of the above holds: it is irrelevant to the
content and the message is in a wrong mental
state.
Evaluation of trust
In this section we conduct evaluation of how human users
lose trust in CASP and his host in case of both content and
mental state relevance failures.
33
expectations, than for shorter, single sentence or phrase
postings by CASP.
Now we compare relevance in Table 1 and failed
relevance in Table 2. Out of hundred posting per user total,
averaging between 2 and 3 for each human user posting,
failures occur in less than 10 different user postings.
Therefore most users did not land to the lost trust. The
friends who were lost due to the loss of trust would be
“inactivated” anyway because of a lack of attention, and
the majority of friends will be retained and “active”.
Fig. 6: Interactive mode of CASP: a user can write an essay
concerning an opinion of her peer.
We measure the failure score of an individual CASP
posting as the number of failures (between zero and two).
In Table 2 we show the results of tolerance of users to the
CASP failures. After a certain number of failures, friends
loose trust and complain, unfriend, shares negative
information about the loss of trust with others and even
encourage other friends to unfriend an friend who is
enabled with CASP. The values in the cell indicate the
average number of postings which failed relevance when
the respective event of lost trust occurs. These posting of
failed relevance occurred within one months of the
experiment run, and we do not access the values for the
relative frequency of occurrences of these postings. On
average, 100 postings were done for each user (1-4 per
seed posting).
One can see that in various domains the scenarios users
loose trust in the system are different. For less informationcritical domains like travel and shopping, tolerance to
failed relevance is relatively high. Conversely, in the
domains taken more seriously, like job related, and with
personal flavor, like personal life, users are more sensitive
to CASP failures and the loss of trust in its various forms
occur faster.
For all domains, tolerance slowly decreases when the
complexity of posting increases. Users’ perception is worse
for longer texts, irrelevant in terms of content or their
Conclusion
We proposed a problem domain of social promotion and
build a conversational agent CASP to act in this domain.
CASP maintains friendship and professional relationship
by automatically posting messages on behalf of its human
host. We demonstrated that a substantial intelligence in
information retrieval, reasoning, and natural languagebased relevance assessment is required to retain trust in
CASP agents by human users.
We performed the evaluation of relevance assessment of
the CASP web mining results and observed that using
generalization of parse thickets for the seed and candidate
message is adequate for posting messages on behalf of
human users.
We also evaluated the scenarios of how trust was lost in
Facebook environment. We confirmed that it happens
rarely enough for CASP agent to improve the social
visibility and maintain more friends for a human host than
34
Dias, J. and Paiva, A. 2005. Feeling and Reasoning: A
computational model for emotional characters. In EPIA Affective
Computing Workshop, Springer.
Domingos P. and Poon, H. 2009. Unsupervised Semantic Parsing,
with, Proceedings of the 2009 Conference on Empirical Methods
in Natural Language Processing, Singapore: ACL.
Galitsky, B., Josep Lluis de la Rosa, Gábor Dobrocsi. 2012.
Inferring the semantic properties of sentences by mining syntactic
parse trees. Data & Knowledge Engineering. Volume 81-82,
November 21-45.
Galitsky, B. 2014. Transfer learning of syntactic structures for
building taxonomies for search engines. Engineering Application
of AI, http://dx.doi.org/10.1016/j.engappai.2013.08.010.
Galitsky, B. 2013. Machine Learning of Syntactic Parse Trees for
Search and Classification of Text. Engineering Application of AI.
Volume 26, Issue 3, 1072–1091.
Galitsky, B. Dmitry Ilvovsky, Sergei O. Kuznetsov, Fedor Strok.
2013. Finding maximal common sub-parse thickets for multisentence search. IJCAI Workshop on Graphs and Knowledge
Representation, IJCAI 2013.
Galitsky, B. 2013. Exhaustive simulation of consecutive mental
states of human agents. Knowledge-Based Systems,
http://dx.doi.org/10.1016/j.knosys.2012.11.001.
Ehrlich H.-C., Rarey M. 2011. Maximum common subgraph
isomorphism algorithms and their applications in molecular
science: review. Wiley Interdisciplinary Reviews: Computational
Molecular Science, vol. 1 (1), pp. 68-79.
Bron, Coen; Kerbosch, Joep. 1973. Algorithm 457: finding all
cliques of an undirected graph", Commun. ACM (ACM) 16 (9):
575–577.
Montaner, M.; Lopez, B.; de la Rosa, J. L. (June 2003). "A
Taxonomy of Recommender Agents on the Internet". Artificial
Intelligence Review 19 (4): 285–330.
Collins, M., and Duffy, N. 2002. Convolution kernels for natural
language. In Proceedings of NIPS, 625–632.
Bhasker, B., K. Srikumar . 2010. Recommender Systems in ECommerce. CUP. ISBN 978-0-07-068067-8.
Trias i Mansilla, A., JL de la Rosa i Esteva. 2012. Asknext: An
Agent Protocol for Social Search. Information Sciences 190, 144–
161.
being without CASP. Hence although some friends lost
trust in CASP, the friendship with most friends was
retained by CASP therefore its overall impact on social
activity is positive.
The robust intelligence is achieved in CASP by
integrating linguistic relevance based on parse thickets and
mental states relevance based on simulation of human
attitudes. As a result, preliminary evaluation showed that
CASP agents are trusted by human users in most cases,
allowing CASPs to successfully conduct social promotion.
The content generation part of CASP can be used at
www.facebook.com/RoughDraftEssay. Given a topic, it
first mines the web to auto build a taxonomy of entities
(Galitsky 2014) which will be used in the future comment
or essay. Then the system searches the web for these
entities to create respective chapters. The resultant
document is delivered as DOCX email attachment.
In the interactive mode, CASP can automatically compile
texts from hundreds of sources to write an essay on the
topic (Fig. 6). If a user wants to share a comprehensive
review, opinion on something, provide a thorough
background, then this interactive mode should be used. As
a result an essay is automatically written on the topic
specified by a user, and published.
References
Abney, S. Parsing by Chunks, Principle-Based Parsing, Kluwer
Academic Pub-lishers, 1991, pp. 257-278.
Cassell, J., Bickmore, T., Campbell, L., Vilhjálmsson, H., and
Yan, H. 2000. Human Conversation as a System Framework:
Designing Embodied Conversational Agents", in Cassell, J. et al.
(eds.), Embodied Conversational Agents, pp. 29-63. Cambridge,
MA: MIT Press.
National Science Foundation. 2013. Robust Intelligence (RI);
www.nsf.gov/funding/pgm_summ.jsp?pims_id=503305&org=IIS
Lawless, W. F., Llinas, J., Mittu, R., & Sofge, D.A., Sibley, C.,
Coyne, J., & Russell, S. 2013. Robust Intelligence (RI) under
uncertainty: Mathematical and conceptual foundations of
autonomous hybrid (human-machine-robot) teams, organizations
and systems. Structure and Dynamics, 6(2).
Hayes G. and Laurel Papworth. 2008. The Future of Social Media
Entertainment. http://www.personalizemedia.com/the-future-ofsocial-media-entertainment-slides/.
Lisetti, CL. 2008. Embodied Conversational Agents for
Psychotherapy. CHI 2008 Workshop on Technology in Mental
Health.
De Rosis, F. Pelachaud, C. Poggi, I., Carofiglio, V. de Carolis, B.
2003. From Greta’s mind to her face: modeling the dynamics of
affective states in a conversational embodied agent, International
Journal of Human-Computer Studies, Vol. 59.
de Salvo Braz, R., Girju, R., Punyakanok, V., Roth, D. and
Sammons, M. 2005 An Inference Model for Semantic Entailment
in Natural Language, Proc AAAI-05
35