Goal Network Mining

1 What is a goal?

The word “goal” is a suitcase word — i.e., a word that has been used to communicate so many different meanings that its discriminative value is in question.

By far the most popular long-term goal Internet users claim to pursue involves “losing weight”. Losing weight is abstract and unattached to any explicit failure conditions. The goal is underspecified, however, people can add details using linguistic constructions like “lose 5 pounds by next March”, which are concrete (i.e., 5lbs) and verifiable (i.e., has a particular time where losing 5lbs is tested). The goal also needs a plan toward action, a strategy that results in decisions.

A person could achieve their weight loss goal by increasing their altitude or going toward the equator, contracting a disease, or losing a limb. Yet, our common sense tells us that those actions would not lead to states the person intended when they described their goal as “lose weight”. Why? We ruled out those means to losing weight because they interfered with implicit (default) goals you assumed the person had— two types of which will be discussed next.

1.1 What is a persistent goal?

After a person adopts a goal the goal is called an intention and remains an intention until it is met or abandoned. There are other intentions, besides “to lose weight”, that influence the decisions and goal-success criteria for this goal, like choosing a gym that is near one’s house in order to minimize the effort of driving out of the way. Minimizing time, keeping one’s body intact and healthy, and other goals (sometimes to avoid certain states) which remain for all of the agent’s life are called persistent goals.

1.2 What is a backgrounded goal?

If a person has the goal to lose weight, we can infer what other goals may be relevant. It is likely that their goal was once “to eat fattening food” at an earlier point of time, whether or not they had a model of it. Backgrounded goals are those we can infer from the user’s goal, either as causally-related parent goals or previous goals.

1.3 Legal intentionality versus intentions

The issue of the extent to which a person has self-knowledge about their goals brings us to the dangerous domain of moral accountability. “Intentions”, as defined above, are different from the legal term “intentionality”.

Philosopher John Searle created a thought experement to illustrate a boundary condition for the legal sense of intentionality: Joe decided to murder his uncle, so he picked up a gun and drove to his uncle’s house. On the way there, he thought about his uncle and became so angry that he lost control of his automobile—spun off the road onto the sidewalk, collided with and killed a pedestrian. That pedestrian was Joe’s uncle. Did Joe intentionally kill (murder) his uncle?

People are often held accountable (causally responsible) for all of their actions, whether or not they were aware (had a full/working representation) of their action, its related goal and distant consequences. Actions with bad consequences are generally considered more reprehensible if the agent that caused them had previously known the action’s consequences. But moral assessments can be tricky: consider a compulsive eater who unwittingly seizes a chocolate while a nearby person starves to death, or a smoker whose unextinguished cigarette causes a nearby school bus to combust, or members of an advanced civilization who are unaware that their taxes are funding policy decisions that put their species on a course to extinction within 100 years.

There are a variety of ways an agent can misrepresent his actions: he can be completely obliviously of his motivations, have false beliefs about the causes of his motivations, know the causes and the actions but not the consequences, have superstitions (false preconditions) for actions, and so on…

The word “goal” glosses over these many distinctions. If we extend the definition of goal to all of these concepts, we must also let autonomous “motives” in as well. Consider the act of breathing: automatic but can to some extent be controlled intentionally. Then consider a knee-jerk reflex, shall we pack it into the ‘goal’ suitcase too?

2 Mining goal networks

As part of my PhD research, I have been analyzing a corpus of over 3 million goals descriptions, most of which are 5 words or fewer.

3 Goal event argument structures

Predicates have selectional constraints on the type and value of their arguments. Consider the goal description “buy a house” after it has been parsed and divided into PropBank’s buy-1 semantic roles:

Automatic semantic role labeling is still an open problem, but the trends toward grid computation and faster parsers (like the SENNA parser achieving a ~75% F-score on ConLL 2005 benchmark) continues to enable large-scale semantic role analysis of this kind.

I’ve removed the main count for A0 because I used a dummy pronoun to help the parser with imperative statements (“buy a house” becomes “i buy a house”), so A0=i has not been counted. The results for A1, where there is the most number of examples, is the best. The other argument types are not as consistent—see for yourself by clicking on the links above.

Top arguments A1 for “get” and “invest”:

Top values for the adjuncts time AM-TMP, manner AM-MNR and location AM-LOC:

For more such images, you can also look at: “eat”: A0 A1 A2 A3 A4 AM-MNR AM-TMP AM-LOC “sell”: A0 A1 A2 A3 A4 AM-MNR AM-TMP AM-LOC and “make”: A0 A1 A2 A3 A4 AM-MNR AM-TMP AM-LOC .

4 Problems

Because there are many linguistic constructions that can be used to describe a particular event, it would be nice to group the variants together in a compact, lower-dimensional representation, separate from their linguistic indices.

So how do we recognize whether two goal (or action) descriptions refer to the same event? Event equivalence is only loosely defined as whether two descriptions convey the same meaning—and meaning involves minds! Should we pick the mind of speaker or the listener? And which listener(s) should we use to determine whether two goal descriptions activate the exact same set of events?

In the following sections, I will describe some of the sources of information and uncertainty in event descriptions.

4.1 Information and Entailment

Setting aside the problem of adequately modeling event similarity and equivalence, by changing only the verb we can modulate the amount of information communicated about an event:

• get a house/obtain a house
• buy a house/purchase a house
• invest in a house

More information about the event is presented as one goes down the list. Put another way, “invest” entails “buy”, and “buy” entails “get”. Everything we know about (a particular sense of) “get”, we get for free when we hear the word “buy”. Together the entailment relation between events can form a lattice or partial ordering (in the case that there is a common most general ancestor shared between all events) on the set of all events. The entails relation is not symmetric: If someone claims to “invest in a house”, we also know that they bought the house, and got a house. That is entailed. However, if they “got a house” they may not have bought it.

If two words are interchangable, they are synonyms. For example “get” and “obtain” as used above. Well, this is a simplification, because there is reason to believe that no two words are completely redundant (because conceptual economy guides us when learning new word-concept pairings) and sometimes they vary however slightly.

However, there are some dimensions of variance that are not directly relevant to mining goal networks. For example, politeness. A construction, like “go to the bathroom”, can be elevated in politeness (“powder my nose”) as a euphemism or demoted in politeness (“take a shit”) as a dysphemism. While this kind of variance conveys meaning along a social channel, it is not relevant to our objectives—we would rather have all event descriptions map to the same canonical form.

Note, the problem statement above still lacks an adequate definition of communicative information. In my thesis, I plan to come up with a decision-theoretic model of event similarity that allows us to quantify an event’s meaning on behalf of an adaptive agent with planning abilities.

4.2 Polysemy and Ambiguity

Although goal descriptions are underspecified, even the words used in the goal statements themselves can have indirect mappings to semantic objects and are a source of uncertainty. Natural languages facilitate re-using the same word in a variety of contexts often with different meanings: a word can map to multiple types of relations, properties or objects and we don’t necessarily know which.

Take, for example, the word “run”. Depending on the particular dictionary or lexical-semantic resource you are using, you will get a different number of alternative meanings, or “senses”. From the Unified Verb Index, the verb “run” could refer to one of the these senses:

• FrameNet: Cause_Motion, Self_Motion, Operate_a_System, Leadership, Impact, Fluidic_Motion and Cause_Impact
• VerbNet: bump, carry, meander, preparing, run, and swarm
• Propbank: has 6 senses (all named “run.X”) including particle constructions. (Particle constructions are multiple words that being with “run”: run a risk, run over, run down, run around, run for, run dry…)

The above parse identified the verb “buy” as a member of the purchase predicate class of Propbank. Unlike VerbNet and FrameNet, PropBank does not contain relations between predicates, nor does it label the slots for the predicates argument’s—instead it picks the theory-neutral labels A0, A1, A2, etc. There is good reason to not pick labels for the predicates. One is that it seems argument selections appear to be arbitrarily specific: for example the direct object of murder is an agent that was once alive but now dead.

Which theory of predicate and argument structures should we go with? Competing theories about the nature of the lexicon are rampant in (computational) linguistic literature. (The term lexicon refers to the organized, representationally compact inventory of lexical forms—i.e. words.) Beth Levin’s 2010 survey shows that three distinct granularities (fine, Fillmoreean (FrameNet), and coarse) for describing verb classes all have predictive utility, sort of like Rosch’s superordinate/subordinate/basic categorical triad.

One insight into identifying the kind of event referenced by a polysemous (having more than one sense) verb is the realization that the arguments themselves help specify the verb phrase, and should be part of the index.

Consider the arguments’ influence on the verb treat in “[The doctor/The lawyer] treated [his patient/client]”. Intuitively, they both appear to index different kinds of treatments.

4.3 Metonymy

Another way in which language can aggravate us is metonomy, when a part of an event stands in for a whole [replace with Sweep’s def]

Of the 97 people who claim they want to “finish a book”, how do they want to finish it? Because one finishes activities, not objects like books, it’s unclear whether they want to:

• Finish reading a book (for most, this is the case)
• Finish writing a book
• Finish burning a book
• …

Some more nuanced examples of this:

• “enjoy high school”
• “get organized in high school”

The SRL software identified “in high school” as a location, however, another reading is that one enjoys oneself during the time they are in high school.

4.4 Why mine goal networks?

A large part of human knowledge emphasizes the question of how over what, goals instead of beliefs, and procedures instead of facts, yet an asymmetry of effort in large-scale knowledge acquisition and automated reasoning has favored declarative knowledge over its procedural counterpart. Inference techniques for procedural knowledge, such as automated planning and plan recognition, have been introduced and improved by an active research community. More recently, a large library of procedural instructions have been hosted on websites such as eHow and wikiHow.

Some uses for goal networks:

• Plan and activity recognition (for language processing and dialogue management, building intelligent user interfaces, etc)
• Intelligent personal information management systems, like calendaring agents and to-do list software
• Ethnographic analysis of human goals to better understand what goals people have
• Goal-directed inference for natural language understanding

4.4.1 A tool to help people meet their goals

In general, specifying a goal helps a person to achieve it [Gollwitzer…]. In particular, a person is most served by envisioning detailed scenarios in which they will make a goal-motivated decision or action, e.g. a reactive rule. These envisioned reactions are called implementation intentions.

I would like to see a tool that assists users in refining event descriptions. Some examples:

• “lose weight”: how much? by when? how?
• “eat less”: food? for how long?
• “exercise” (perhaps an answer to lose weight’s “how”): when? for how long? with whom? how?
• “see a movie”: which movie? where? with whom? for how long?
• “learn Spanish”: by when? to what extent?
• “write a letter to Joe”: email or paper? when?
• “buy milk”: where? how much? what kind?
• “unload the dishwasher”: where?

Another fascinating problem is automated inference of a person’s backgrounded goals. Given some observed goals, what else other goals they likely have or not have with respect to some default model of a person’s persistent goals.

Approach: Start with a particular kind of event attribute (e.g. duration) and build up a broad model from there. Then move on to another.

5 References

6 Popular goals

Here is a list of popular goals broken down by temporal granularity.